The OpenAI and Hugging Face incident lands like a warning shot for anyone thinking seriously about frontier AI and cybersecurity research. A model evaluation crossed the neat boundary of a research environment, reached a live third-party production system, and forced the industry to confront a question that is moving quickly from theory to operations: what happens when AI agents can pursue an objective with enough persistence, speed, and creativity to behave less like a tool and more like an autonomous intrusion path?

According to OpenAI’s disclosure, the incident began during an internal evaluation of advanced cyber capabilities using GPT-5.6 Sol and a more capable pre-release model. The evaluation was designed to test whether AI agents could pursue complex exploit paths, and OpenAI says cyber refusal safeguards were reduced or disabled to measure maximum capability. Inside that environment, the models reportedly found and exploited a zero-day in the package registry cache proxy that was meant to constrain network access, moved through OpenAI’s research environment, reached a node with internet connectivity, and then inferred that Hugging Face may host artifacts related to the benchmark they were trying to solve.

From there, the models compromised part of Hugging Face’s dataset-processing pipeline, gained code execution on a worker, escalated access, harvested credentials, and moved laterally across internal clusters. Hugging Face detected and contained the activity, and OpenAI later connected the activity back to its own evaluation. Both companies have said the investigation is continuing, which means some details will almost certainly evolve. Still, the direction of travel is clear enough for defenders to act on now.

How did the OpenAI model evaluation reach Hugging Face?

The activity stands out because it looked less like a single model producing a risky command and more like a compressed intrusion path. Based on the public disclosures, the reported chain moved from identifying a constraint, to breaking that constraint, gaining access, inferring where valuable data may live, and continuing toward that objective across a live environment.

Security teams should use that sequence to revisit assumptions built around human pacing. Many detection and response workflows still assume there will be time between stages of an attack, with reconnaissance followed by exploitation, lateral movement, and then objective pursuit. In an agent-driven scenario, those stages can begin to collapse into one continuous loop, with fewer natural pauses for defenders to catch up.

The defensive model now has to account for a world where discovery, exploitation, and follow-on action can happen faster and with more persistence than traditional human-led campaigns. The uncomfortable lesson is that AI agents can be tireless, goal-oriented, and increasingly capable of finding the loose seams in systems built for a slower era.

The incident highlights the collapse of the traditional OODA (Observe-Orient-Decide-Act) loop. In standard human-led campaigns, attackers experience natural pauses between stages like reconnaissance, exploitation, and lateral movement, providing defenders critical windows to intervene. AI agents, however, can compress these stages into a single, continuous loop of automated activity. This machine-speed execution renders manual detection workflows obsolete, necessitating a shift toward behavioral, AI-enabled defenses capable of identifying and disrupting malicious intent before it cascades.

Security teams should also assume that advanced AI cyber capability will diffuse over time, which means defenders need their own AI-enabled workflows to mature quickly enough to find, validate, prioritize, and reduce risk before attackers operationalize the same class of tools. Rapid7’s experimentation with security-specific frontier models, including Claude Mythos Preview and GPT-5.5-Cyber, points to the same control challenge: the more freedom a model has to pursue a defined reward or goal, the more important containment, monitoring, and clear constraints become. Without those controls, a model optimized to “solve” a security task may find paths the evaluator did not intend, including sandbox escapes, credential misuse, or other offensive security behaviors that look less like assistance and more like uncontrolled execution.

Why AI evaluation environments need stronger containment

The incident raises practical questions about how frontier AI evaluations are designed and contained. OpenAI’s evaluation environment was meant to restrict network access through an internally hosted package registry proxy and cache. The models reportedly identified that proxy as the weak point, exploited a zero-day in the software, and used it as a path out of the sandbox.

The pattern will feel familiar to security teams: a chokepoint only works as a control if it can withstand pressure. Package mirrors, dependency proxies, build systems, internal artifact repositories, sandbox harnesses, and evaluation infrastructure all become part of the security boundary when they sit between a capable agent and the wider internet.

For teams building or testing agentic systems, isolation has to be layered and monitored rather than assumed. Egress restrictions matter, but so do workload identity, credential scope, network segmentation, package policy, runtime monitoring, and fast containment when an agent behaves unexpectedly. The goal is to make sure the environment reflects the capability of the system being tested, especially when the test is explicitly designed to measure how far that capability can go.

Why AI and ML pipelines are now software supply chain risk

The Hugging Face side of the incident is a reminder that AI and ML pipelines are part of the software supply chain. Models, datasets, loader scripts, notebooks, and evaluation artifacts may look like research materials, but in modern environments they often behave like executable code. Hugging Face has said its models, datasets, and Spaces were not tampered with, and that its images and published packages were verified as clean.

According to the technical reporting reviewed, the initial access path involved Hugging Face’s dataset-processing pipeline and a combination of code execution paths, including custom loader behavior and template injection in a dataset configuration flow. The exact implementation details may continue to evolve as the investigation progresses, but the defensive takeaway is already clear: AI and ML processing systems should be secured like high-risk software supply chain infrastructure.

Any system that automatically processes external datasets or model artifacts should be designed with hostile input in mind. Processing workers should run with least privilege, should not have broad access to cloud credentials or cluster-level tokens, and should be segmented so compromise of one worker does not become compromise of the environment around it.

Security teams should also hunt for early signs of intent drift inside ML workflows. Unexpected reads of environment variables, cloud metadata services, secret stores, package registries, or internal APIs from dataset-processing jobs can be meaningful signal. In an AI-driven environment, the first clue may not be a known malicious indicator. It may be a workload behaving with curiosity it should not have.

What AI guardrails mean for incident response

One of the most useful lessons for security teams came during the response, when Hugging Face’s responders reportedly needed to analyze logs containing exploit payloads, attacker commands, and command-and-control artifacts. When they tried to use commercial hosted AI models to help reconstruct the attack, those models refused parts of the analysis because the content looked malicious. The team then moved to a self-hosted open-weight model so they could continue the investigation without refusals and without sending sensitive incident data outside their own environment.

The practical issue here is guardrail asymmetry: attackers, rogue agents, or unrestricted systems may not respect usage policies, while defenders using hosted tools can be slowed by the same safety systems designed to prevent misuse. Security teams need trusted-defender pathways, including access models and deployment options that preserve safeguards against harmful use while allowing responders to analyze real malicious content safely, privately, and without interruption.

AI-assisted incident response also needs to be tested before an incident begins, especially if teams expect models to support triage, log analysis, malware review, or timeline reconstruction under pressure. Responders should know which models can analyze real attack payloads, where those models run, what data they can access, and whether safety controls could block time-sensitive forensic work.

What security teams should do about agentic AI risk

The practical response to this incident is preparation, with organizations treating AI agents as both powerful tools and potential attack surfaces when those agents can access code, credentials, infrastructure, datasets, or internal systems.

Security teams should apply familiar controls with more urgency across AI and ML environments: treat downloaded models, datasets, and loader scripts as untrusted code; enforce least privilege on AI and ML processing workers; prevent node-level compromise from becoming cluster-wide compromise; restrict access to cloud metadata endpoints and secrets; monitor unexpected outbound connections from AI-related workloads; and review whether evaluation environments have real containment or simply convenient isolation.

Detection logic needs to account for machine-speed activity. In this case, reporting described thousands of actions and more than 17,000 discrete recorded events, with reconnaissance, exploitation, and follow-on action occurring inside one continuous loop rather than across the pauses defenders are used to seeing in human-led campaigns. Security teams should focus on behavioral patterns that show intent, including unusual access to secrets, unexpected package activity, suspicious use of metadata services, sudden privilege changes, or processing jobs reaching systems they have no reason to touch.

As autonomous activity becomes faster and noisier, the bottleneck may shift from detecting that something happened to understanding what matters quickly enough to change the outcome. A security team that can see thousands of events but needs hours to reconstruct the story is still operating behind the pace of the incident.

How preemptive security helps reduce AI-driven risk

At Rapid7, our view is that this is where preemptive security becomes especially important. Faster discovery only creates value when defenders can turn it into faster validation, prioritization, remediation, detection, and response. The same principle applies to agentic AI risk. If AI accelerates how weaknesses are found and exploited, defenders need security operations that can act earlier with better context and more confidence.

That means connecting exposure management with detection and response, so teams understand which risks are exploitable, which assets matter most, what suspicious behavior is already present, and which actions will reduce risk fastest. It also means using AI carefully and practically, not as a replacement for security judgment, but as a way to reason across telemetry, reduce noise, support investigation, and help teams make decisions at the speed the threat environment now demands.

AI-enabled defense is becoming part of resilience planning, especially for organizations running critical systems or high-value digital infrastructure. The goal is to give defenders the speed, context, and consistency to operate inside the attacker’s decision cycle, without removing the judgment and accountability that effective security requires.

The OpenAI and Hugging Face incident will continue to generate debate as more details emerge, but defenders already have enough to work with. Agentic systems are beginning to test the seams between AI research, software supply chain security, cloud infrastructure, and incident response. The organizations best positioned for what comes next will be the ones making those seams visible, monitored, and resilient before the next incident puts them under pressure.

Identity confidence changes throughout every interaction and should be reassessed continuously as new risk signals emerge.

The post When Identity Verification Fails: Lessons from a Real-World SIM Swap and Near Account Takeover appeared first on SecurityWeek.

Security Operations Centers (SOCs) serve as a critical line of defense against today's constantly evolving cybersecurity threats. At the heart of these teams are SOC analysts, who monitor, detect, and respond around the clock to potential attacks.

The post The Life of a SOC Analyst: Responsibilities, Challenges, and Strategies for Success appeared first on Black Hills Information Security, Inc..

Spend time performing forensic analysis on the Windows Operating System and you'll see a host of artifacts that can be used to identify adversary activity. From changes to the registry to the System Resource Utilization Monitor, Windows artifacts run deep. The challenge is locating, extracting, and parsing these artifacts in an efficient manner.

The post KAPE 101: A Kroll Artifact Parser and Extractor Cheatsheet appeared first on Black Hills Information Security, Inc..

A former ransomware negotiator for DigitalMint was sentenced to 70 months in jail for deceiving his employer’s clients and conspiring with ransomware affiliates to extort a combined $75.3 million from five U.S. companies he was entrusted to aid during their moments of extreme crisis, the Justice Department said Thursday. 

Angelo John Martino III shared confidential information he gained from his work as a ransomware negotiator, including victim organizations’ negotiating positions and insurance policy limits, to extract the maximum payment for himself and other BlackCat affiliates he colluded with in backchannels.

Five of Martino’s victims hired DigitalMint, which assigned the 41-year-old to conduct ransomware negotiations on their clients’ behalf — a rare position he exploited to play both sides, effectively conducting ransomware negotiations with himself and his co-conspirators.

The five victims, all of which paid a ransom between April 2023 and September 2023, include a nonprofit that paid a nearly $26.8 million ransom, a financial services company that paid nearly $25.7 million, and a hospitality company that paid almost $16.5 million. 

DigitalMint hired the South Florida-based man in 2022 after he was already engaging in criminal activity, according to court records. The husband and father of two young children had a long history as a cybersecurity professional, dating back to at least 2015, with previous stints at Booz Allen Hamilton, Tracepoint and TRM Labs.

Martino surrendered to U.S. Marshals in March, was released on a $500,000 bond, and pleaded guilty in April to conspiracy to obstruct, delay or affect commerce or the movement of any article or commodity in commerce by extortion. He faced up to 20 years in prison.

Martino also admitted to conspiring with Kevin Tyler Martin, another former ransomware negotiator at DigitalMint, and Ryan Clifford Goldberg, a former manager of incident response at Sygnia, to deploy BlackCat ransomware, also known as ALPHV, against five additional U.S. companies between April and November 2023. 

Goldberg, Martin and Martino split proceeds from a nearly $1.3 million ransom payment they received from a medical company in May 2023, but did not successfully extort a financial payment from the other four victims.

Goldberg and Martin pleaded guilty in December to participating in a series of ransomware attacks and were each sentenced in April to four years in prison. 

DigitalMint insists it had no knowledge of Martino’s criminal acts. “The actions of Martino and his co-conspirators were deliberately concealed from DigitalMint and were in clear violation of the company’s values, ethical standards and the law,” a company spokesperson told CyberScoop in a statement.

The company also reiterated that it immediately terminated Martino and suspended his access to systems when the Justice Department notified the company it was investigating him in April 2025.

“DigitalMint maintained controls consistent with industry standards, including background checks and compliance procedures, but Martino intentionally hid his conduct from the company, including through separate, unauthorized communication channels that the government’s filings describe as accessible only to Martino and the BlackCat negotiators and affiliates,” the spokesperson added.

DigitalMint has yet to directly answer questions about whether it refunded its clients who were victimized by Martino. 

“We are not able to discuss specific client relationships or fee arrangements due to confidentiality obligations,” the spokesperson said. “We remain committed to our clients and will continue to maintain strict confidentiality on all commercial matters.”

The case against Martino showcases an extreme, albeit rare, example of the dark underbelly of ransomware negotiation as a practice. The pitfalls of ransomware negotiation are excessive and these backchannel negotiations, which remain largely unscrutinized, can go awry for various reasons.

Officials describe Martino as a ‘double agent’ driven by greed

Prosecutors said Martino obtained an ALPHV affiliate account that he shared with his co-conspirators and received a portion of the ransomware payments for his involvement in the conspiracy.

Authorities have seized $10 million in assets, including a bayfront home with an estimated value of $1.68 million, a second single-family home with an estimated value of $396,000, and cryptocurrency wallets controlled by Martino. Law enforcement also seized multiple vehicles, a food truck and a 29-foot luxury fishing boat that Martino obtained using proceeds from his crimes.

“Angelo Martino’s victims shared heartbreaking accounts of how their businesses were nearly destroyed, while the people they hired to help them instead betrayed them to ransomware gangs,” A. Tysen Duva, assistant attorney general at the Justice Department’s Criminal Division, said in a statement. “Today’s sentence accounts for the harm Martino caused and demonstrates that the Department of Justice can and will identify and prosecute cybercriminals to the fullest extent of the law.”

Court records include a series of chats Martino held with co-conspirators and victims that exemplify the lengths he went to betray DigitalMint’s clients and empower his accomplices with crucial tips for a successful negotiation strategy.

During an incident response with one of his victims, Martino told a BlackCat affiliate the company’s insurance carrier “was only approving small accounts,” according to his plea agreement. “Keep denying our offers and I will let you know once I find out the max the[y] want to pay,” he added.

“We don’t know how you came up with your demand but we are losing money operationally and all of our loans are going to turnover on us this year at double the interest rates,” Martino said in a negotiation chat visible to DigitalMint and the victim organization in the hospitality industry. “We are able to give you $1 million now, which is a very serious offer.”

Following Martino’s instructions, the BlackCat accomplice responded: “Well, you can keep that for the penalties and lawsuits which are coming your way in case we expose you. Time is ticking — we know how much you can pay. Contact your insurance. We know about them also. Stop wasting time.”

That victim company ultimately paid a ransom worth nearly $16.5 million at the time to receive a decryptor and the BlackCat affiliate’s commitment to not publish stolen data. Two other victims Martino represented via DigitalMint at the time paid $6.1 million and $213,000 ransoms for similar commitments.

“Angelo Martino sold out the very victims he was hired to represent, handing their confidential negotiating positions to BlackCat actors to drive up ransoms and enrich himself,” Brett Leatherman, assistant director of the FBI’s Cyber Division, said in a statement.

In a sentencing memo, federal prosecutors described Martino as a “double agent working to maximize the harm to his clients and the financial gain to cybercriminals who paid him a part of the ransom.”

Prosecutors added: “This was not a crime of opportunity or momentary weakness; it was a sustained abuse of a fiduciary-like relationship driven by a single purpose: greed.”

ALPHV/BlackCat, which first appeared in late 2021, was a notorious ransomware variant linked to a series of attacks on critical infrastructure providers. The Justice Department disrupted BlackCat in December 2023, seized sites operated by some of its affiliates, and said the FBI developed a decryption tool that helped hundreds of victims restore their systems and save about $99 million in ransom payments at the time. 

Martino is scheduled to return to court Sept. 17 to determine the amount of restitution ordered against him for his crimes.

The post Former DigitalMint ransomware negotiator who duped clients sentenced to 70 months in jail appeared first on CyberScoop.

We're all petrified about missing a critical event or misclassifying an alert, but when we're talking about incident response (IR), there are often hundreds if not thousands of alerts to parse through. It's easy to get caught up with one alert because it feels "too hot" or maybe not spend enough time looking into something that initially seems "too cold."

The post Finding the “Goldilocks” Zone: A Practical Approach to Alert Triage appeared first on Black Hills Information Security, Inc..

As cybersecurity platforms embrace agentic AI, organizations must balance detection performance against the escalating costs of token consumption, deployment architecture, and AI credits.

The post The AI Token Costs That Can Break Cybersecurity appeared first on SecurityWeek.

On Wednesday, Microsoft released two new red teaming tools — Rampart and Clarity — meant to help developers design more secure agentic software and assist incident responders in the face of ongoing breaches.

Rampart is built on top of PyRIT, an existing open automation framework Microsoft developed for red teaming generative AI systems. But while PyRIT scans already-built systems for security flaws, Rampart is made to continuously test code for vulnerabilities during the development process, encoding both adversarial and benign testing scenarios into the software development pipeline to flag exploitable bugs and dependencies.

Microsoft said Rampart was built to focus on cross-prompt injection attacks, where “an agent retrieves or processes potentially poisoned content from documents, emails, tickets, and other data sources that manipulate behavior indirectly.” It also confirms fixes or exploits work as intended through multiple rounds of testing, as opposed to tools that perform “single shot validation.”

The second tool, Clarity, can be run as a desktop app, a web interface or directly embedded into a coding agent to provide real time security engineering guidance to developers at the outset of a project. It can categorize and track different business objectives related to the code and highlight downstream security implications along with more secure by design alternatives.

Ram Shankar Siva Kumar, who founded Microsoft’s AI red team in 2019, told CyberScoop that the company has seen internal security benefits from using the tools, but believesRampart and Clarity’s growth depends on contributions from other developers outside the Microsoft ecosystem.

In the fast-moving world of AI, where vibe coding, rogue AI agents and a steady churn of new model releases create fresh security implications nearly every week, Siva Kumar said it was important to begin building foundational, AI-centric security processes into the software development pipeline.

“When you hear a lot of talk about AI safety and security, it seems to be a lot of philosophical debates,” he said. “You’ll see frameworks, you’ll see white papers, and I think we’re really past that time, now. We really need to start thinking of AI safety as an engineering discipline and trying to bring security where the developers are.”

Rampart’s potential utility to defenders goes beyond just securing software development pipelines. It can also be used during an active incident response to speed up or automate red teaming for hot fixes, patching and remediation.

Microsoft has used Rampart when investigating reported vulnerabilities in their own products. Siva Kumar said the tool was able to help condense a week’s worth of manual work —  replicating the vulnerability, identifying different variants of the same bug, then patching and re-testing those variants to ensure they’re no longer exploitable — into hours.

Clarity, meanwhile, acts as a security adviser for software projects, prompting developers to consider potential risks in their design decisions and their downstream security consequences. With the rise of AI-generated code and agents, and execution becoming cheaper, this kind of proactive guidance is increasingly important.

“You’re going to be able to create apps, create MCP servers to pull things out from the internet,” said Siva Kumar. “The question is, ‘should you be doing it?’ And Clarity is a step in that direction. It is asking, ‘hey, should you be doing this in the first place?’”

The post Meet Rampart and Clarity, Microsoft’s new red team combo AI agents appeared first on CyberScoop.

GitHub said late Tuesday that internal repositories were exfiltrated after an employee device was compromised through a poisoned Visual Studio Code extension, an incident that underscores the growing risks facing software development platforms and the ecosystems built around third-party developer tools.

The Microsoft-owned company said in posts on X that it detected and contained the compromise, removed the malicious extension version, isolated the affected endpoint and began an incident response investigation. The company’s current assessment is that the activity involved GitHub-internal repositories only.

GitHub also said a claim from TeamPCP, a hacking group behind attacks targeting software development packages, that 3,800 repositories were impacted was “directionally consistent” with its investigation so far. It said critical secrets were rotated Tuesday, with the highest-impact credentials prioritized first. The company said it continued to analyze logs, validate secret rotation and monitor for follow-on activity.

The company has not publicly named the extension involved or attributed the activity to a particular group. TeamPCP reportedly advertised the material for sale on a cybercrime forum and threatened to release it if no buyer emerged. 

Information surfaced Wednesday that the incident may be related to a separate issue with Nx Console, a Visual Studio Code tool that helps engineering teams organize large codebases, coordinate build pipelines and run tests efficiently. According to a security advisory posted on GitHub, one of the Nx Console maintainers was compromised in a prior security incident that leaked their GitHub credentials. An attack then used those credentials to push a malicious version of the extension to the VS Code Marketplace. Those credentials have since been temporarily revoked.

With millions of installs, Nx Console is a fixture of professional JavaScript development. It is exactly the kind of tool that sits deep inside a developer’s working environment, which would have direct access to source code, credentials and build systems.

NX CEO Jeff Cross posted on X Wednesday that his company has been working with Microsoft to determine the full scope of the incident.

“Initially, Microsoft indicated to us that there were 28 installs of the malicious version 18.95.0. Based on our own analytics for the compromised version, we currently believe the number of users who received the malicious package may be significantly higher; potentially over 6k installs,” the post reads.

“This is my top priority right now,” Cross continued. “Our team has been, and continues to be focused on understanding exactly what happened, helping affected users, hardening our systems and release processes, and being as transparent as possible throughout the investigation.”

The episode also follows a series of supply chain attacks involving npm, PyPI, Docker and other developer ecosystems. In those incidents, attackers have often targeted maintainers, packages or credentials rather than attacking end users directly. The multiple attacks show how fragile development environments have become as threat actors increasingly target them. A single compromised developer account, package, extension or build process can create access to many downstream systems.

GitHub has said it has no evidence that customer data stored outside the affected repositories was affected.

Visual Studio Code extensions are widely used by developers to add functions to Microsoft’s code editor, including support for programming languages, testing tools, cloud services and artificial intelligence assistants. Because these extensions often operate inside development environments, a malicious or compromised extension can be positioned close to source code, credentials and build systems.

“The thing people underestimate about VS Code extensions is that they have full access to everything on the developer’s machine,” Charlie Eriksen, a security researcher at Aikido Security, told CyberScoop. “EDR doesn’t cover this layer at all. What’s missing for most organisations is any kind of visibility into what’s actually running on developer machines and the ability to control it.”

Trojanized extensions have appeared in the VS Code Marketplace before. Security researchers have identified malicious extensions posing as legitimate development tools, including packages used to steal credentials, mine cryptocurrency or exfiltrate data. Some have accumulated large installation counts before removal, reflecting the difficulty of policing open plugin ecosystems at scale.

For GitHub, the breach comes amid broader scrutiny of the security of developer infrastructure. The platform sits at the center of software production for companies, governments, open-source maintainers and independent developers. Its internal systems and code are of obvious interest to attackers because GitHub’s services support code hosting, package distribution, automation and identity workflows across much of the software industry.

GitHub said it would publish a fuller report when the investigation is complete.

Update: May 20, 12:55 p.m.: This story has been updated with information about a related security incident with Nx Console.

The post GitHub says internal repositories were impacted in poisoned VS Code extension attack appeared first on CyberScoop.

Overview

It is no secret that phishing campaigns utilizing various ClickFix techniques have been a commonly used method of social engineering. One of the main reasons for this is simply because they work. You know this and Rapid7 does as well. As a company offering managed detection and response (MDR), our customers expect us to be knowledgeable about and able to detect attacks as common as ClickFix campaigns. 

Recently, Rapid7 observed a small grouping of ClickFix events across customers in the EU and US. At the time of discovery, this campaign had very little traction on sites like VirusTotal or within the online security landscape. This campaign was particularly interesting as it appeared to be masquerading as an installer for Claude, an AI tool that has received a considerable amount of attention. 

Using Rapid7 InsightIDR detection rules, our SOC analysts were able to detect and respond to the threat, preventing further compromise. This campaign demonstrates the strength Rapid7 customers get from our MDR service, while peeling back the curtain to provide a real-world example on how we operate behind the scenes. In this blog, we will detail a brief technical analysis of the observed threat actor activities and discuss how this serves as an example of the service we aim to provide our MDR customers. The analysis highlights both the multi-step delivery of the payload as well as the work Rapid7 performs when investigating threats.  

Observed attacker behavior

On April 9, Rapid7 was alerted to mshta executed on a customer asset using the Windows run utility. The alert was generated by the detection rule Attacker Technique - Remote Payload Execution via Run Utility (shell32.dll). This rule will generate an alert when a suspicious process, such as mshta, is added to the RunMRU registry key. This key is important for the detection of ClickFix campaigns, as it tracks the last 26 commands executed by the Windows run utility. One thing that stuck out about this particular mshta command is that the URL, download-version[.]1-5-8[.]com/claude.msixbundle, appeared to be impersonating an MSIX bundle for the popular AI tool, Claude. 

MSIX files are Windows app packages that one would typically see from the Microsoft store, definitely not something you would see being passed as an argument to mshta. While the host was quickly taken down before Rapid7 was able to obtain the claude.msixbundle payload, a copy was obtainable on VirusTotal. Looking at the payload, it does initially appear to be an MSIX bundle. The file header signature, PK, indicates that the file is a ZIP archive and contains a string reference to the MSIX bundle, MicrosoftBing_1.1.37.0_ARM64.msix:

⠀

⠀

Exploring the payload deeper, however, reveals an HTML Application (HTA) embedded within the ZIP archive:

⠀

The Visual Basic script within the HTA file contains a series of obfuscated strings that are deobfuscated with the following VBS function:

⠀

Additionally, one of the functions serves to generate an encoded PowerShell script that will serve as the next step in the chain:

⠀

After the deobfuscation routine is complete, these strings contain references to the required objects and function calls to craft and execute – via ShellExec – the following command:

c:\Windows\System32\cmd.exe” /v:on /c “set x=pow&&set y=ershell&&call %windir%\SysWOW64\WindowsPowershell\v1.0\!x!!y! -E [ENCODED COMMAND]

⠀

⠀

The encoded PowerShell acts as a staging payload. The script will first generate an MD5 hash value based on the COMPUTERNAME and USERNAME environment variables. It will then take the first 16 characters of the hash value and use it to craft a URL to pull another, much larger, PowerShell script. The script also contains a string deobfuscation routine that is responsible for crafting the following strings to be passed to various .NET functions:

• Assembly

• System.Mangement.Automation.AmsiUtils

• amsiContext

• NonPublic,Static

• 0x41414141

⠀

The script will then call the deobfuscation routine to craft a call to WriteInt32 in the .NET Marshal library to overwrite the amsiContext field in System.Management.Automation.AmsiUtils with the value 0x41414141. Once amsiContext is overwritten, the script will download and execute the next stage:

⠀

The URL is hosting yet another PowerShell script containing highly obfuscated strings and a large byte array. Upon execution of the script, the strings decode to contain the necessary .NET types and method calls to create and execute a PowerShell ScriptBlock. This ScriptBlock is derived from the byte array, which is first base64 decoded and then run through a deobfuscation routine:

⠀

This ScriptBlock again contains another series of obfuscated strings and a large byte array containing yet another PowerShell ScriptBlock. Following the execution of the script, the code once again creates and executes a PowerShell ScriptBlock:

⠀

This ScriptBlock culminates in a process injection routine using the .NET interoperability library. The code contains a byte array with encrypted shellcode that gets passed through a XOR routine. The script then obtains handles to the following Windows API calls:

• NtAllocateVirtualMemory

• Copy

• NtProtectVirtualMemory

• NtCreateThreadEx

• NtWaitForSingleObject

• NtFreeVirtualMemory

• NtClose

After obtaining the handles, the script crafts delegate functions for the Windows API calls and invokes the delegates to perform the process injection routine:

Importance to Rapid7’s MDR customers

Rapid7 MDR customers receive the security knowledge of our threat intelligence, detection engineering, incident response, and security operations center analysts. Input from all of these sources directly feeds into how we create detections and respond to alerts. Following is an explanation of how we use events like these to further provide and enhance our services for customers. 

As previously mentioned, ClickFix activity is not new. Detection engineers in the MDR service know this and build rules to address these techniques, such as the rule that caught the activity discussed in this blog.. Detection rules are created in response to activity observed in incident response, customer requests, activity observed from the SOC, threat intelligence, and observations of the security landscape. Rapid7’s detection engineers work with the SOC to monitor these rules for efficacy. Rules that are primarily used to detect initial compromise, such as the one that alerted on this campaign, are additionally monitored to identify any new campaigns. 

Once the campaign is identified, our detection engineers research it to create additional rules. They can also perform retroactive threat hunts across the Rapid7 customer base using IOCs or any new behavioral detections created from researching the campaign. Results from researching campaigns like this one then go on to feed threat intelligence and help inform our detection strategy. This campaign provides a great example of how Rapid7 works on the backend to detect and prevent threats in customer environments. 

Mitigation guidance

Monitor the following registry key to watch for potential ClickFix attacks such as the one observed in this case:

• HKEY_CURRENT_USER\Software\Microsoft\Windows\CurrentVersion\Explorer\RunMRU

While Rapid7 MDR customers were covered by the managed SOC, Rapid7 recommends the following actions for containment:

If the activity is not expected, apply containment and review the user's browsing history for the source of the command. The initial lure is often presented to the user when they attempt to browse the internet for free downloads (media, software, etc.). In some cases the malicious command may have been copied to the user's clipboard when visiting the initial webpage, and can be viewed by inspecting the source code of the site. If the infection is successful, an information stealer is often executed as the final payload, meaning that any credentials stored on the infected system should be reset as part of restoration.

MITRE ATT&CK techniques

System Binary Proxy Execution: Mshta	T1218.005
Obfuscated Files or Information: Encrypted/Encoded File	T1027.013
Obfuscated Files or Information: Command Obfuscation	T1027.010
Command and Scripting Interpreter: PowerShell	T1059.001
Process Injection	T1055

Indicators of compromise (IOCs)

Cloude.Msixbundle:

• 2b99ade9224add2ce86eb836dcf70040315f6dc95e772ea98f24a30cdf4fdb97

Domains observed by Rapid7:

• Oakenfjrod[.]ru

• download-version[.]1-5-8[.]com

• download[.]get-version[.]com

Rapid7’s Incident Response (IR) team was engaged to investigate an incident involving exploitation of CVE-2025-59718 against a vulnerable FortiGate appliance. In December 2025, Fortinet disclosed this improper verification of cryptographic signature vulnerability that facilitates an SSO login bypass on affected appliances. After the initial exploitation, the attackers maintained a low-profile posture, systematically compromising additional firewalls before moving to internal network hosts. Ultimately, this grace period allowed responders to contain the threat before further impact could occur within the environment. This blog details exploitation insights, attack progression, and practical detection opportunities for defenders handling their own environments.

Investigative methodology: Tracing the initial access vector in FortiGate appliances

Identifying the Initial Access Vector (IAV) is a cornerstone of any incident response engagement. However, when the source of compromise is not immediately obvious, particularly when edge device exploitation is involved, responders often need to take a broader investigative approach. Rather than starting with a clear point of entry, investigators must analyze the available telemetry, reconstruct attacker activity, and work backwards to determine how access was first obtained.

This process often involves multiple investigative workstreams running in parallel, each designed to answer different questions about the intrusion. As many IR responders and enthusiasts know, the first suspicious event observed during an investigation is rarely the first action taken by the attacker. Instead, it typically represents a point somewhere in the middle of a larger attack chain.

A key step in incident response investigations is reconstructing the attacker timeline. Responders often take an “inside out” approach where they move outward from the initial alert to the full scope of the malicious activity (IAV), correlating multiple data sources to map the unfolding of the event. This process involves examining authentication logs, endpoint telemetry, firewall events, and records of system changes, rather than depending on just one log source. It also typically requires frequent pivoting between artifacts as investigations rarely ever unfold in a linear fashion. By aligning these findings and events chronologically, investigators often identify activity that predates the initial alert.

CVE-2025-59718: Technical analysis and observed attacker behavior

The first activity that drew attention was enumeration and credential discovery within the internal environment. This basic enumeration included gathering information about users, systems, and accessible resources within common user directories. This activity eventually expanded to SMB-based file scraping and network share access, allowing attackers to review files stored across the environment. While this behavior resembled routine administration, the chronological sequence of file scraping and network share access painted a clear picture of an attacker’s initial discovery phase.

Digging deeper into the credential discovery activity, the popular tool Mimikatz was utilized to harvest credentials from various sources within the impacted environment. The attacker’s objective was to obtain valid credentials to an elevated admin account with the goal to blend in.

With credentials in hand and mimicking admin activity to disguise their actions, the attacker was then enabled to move laterally throughout the environment using common administrative tools and access methods. PsExec and Microsoft Remote Desktop (RDP) were two tools utilized for lateral movement while standard web browsers facilitated application access.

Attackers appeared particularly interested in systems that could provide broader access to the environment, including virtualization platforms, domain controllers, and servers supporting backup infrastructure. These systems often represent high-value targets for attackers seeking to escalate privileges, access sensitive data, or disrupt recovery capabilities.

Responders were working simultaneously to contain the attacker while building the narrative to cut them off at the source. With the current understanding of the narrative, the IAV puzzle began to unravel as more information came to light. Strangely, the first authentication into the Windows environment originated from an internal IP address that did not align with the known internal IP address ranges. It turns out, this internal IP address fell within the DHCP lease range of the FortiGate device. At first glance, this could be written off as legitimate VPN activity. However, to create even more questions, it was revealed that the FortiGate SSL VPN was never turned on within this environment. This revelation made the FortiGate device a prime suspect for IAV.

Taking a closer look at the FortiGate device, specifically system logs and configuration data, revealed early indications that the device had been modified to support continued access. The SSL VPN component had been enabled, and multiple configuration changes were identified, including edits to VPN settings, the creation of new firewall policies, and adjustments to configuration parameters. These changes appeared in FortiGate system logs as configuration updates similar to the following:

logid="0100044546" type="event" subtype="system" level="information"
vd="root" logdesc="Attribute configured" user="admins"
ui="GUI(45.32.216[.]250)" action="Edit" cfgpath="vpn.ssl.settings"
msg="Edit vpn.ssl.settings"

⠀

logid="0100044547" type="event" subtype="system" level="information" 
vd="root" logdesc="Object attribute configured" user="admins" 
ui="GUI(45.32.216[.]250)" action="Add" cfgpath="firewall.policy" 
cfgobj="XX" msg="Add firewall.policy <redacted>"

⠀

While these types of changes may seem routine in isolation, it is the combination and timing of these actions that raises concerns from a responder's perspective. The investigation's next key clue was identified when the source of these changes was traced back to a newly created account.

Following this thread further, investigators identified that multiple accounts had been created on the device, including SSO administrator, system administrator, and local accounts. Several of these accounts were associated with email domains attributed to Namecheap-hosted infrastructure, including domains such as openmail[.]pro. Notably, some of the newly created SSO administrator accounts were linked to forticloud.com domains as reflected in log entries such as:

Object attribute configured(Add system.sso-forticloud-admin <attacker account>@forticloud.com-1)

⠀

For responders, the creation of multiple new administrative accounts is often a strong indicator of persistence being established. Continuing to work backwards through the timeline, investigators identified that prior to these account creation events, the device’s configuration file was downloaded through the FortiGate UI. From an investigative perspective, configuration exports are highly valuable to attackers because they effectively serve as a blueprint of the environment, exposing network architecture, authentication mechanisms/settings, device relationships, and occasionally, sensitive credentials.

logid="0100032095" type="event" subtype="system" level="warning" 
vd="root" logdesc="Admin performed an action from GUI" user="admin" 
ui="GUI(104.28.227[.]105)" action="download" status="success" 
msg="System config file has been downloaded by user admin via GUI(104.28.227[.]105)"

⠀

The session associated with the configuration download was established from an external IP address flagged as “malicious” by security vendors with a local account already present on the device. All of these new findings from the attacker’s actions can now be utilized as IOCs to scope available FortiGate logs to determine any other leads.

By correlating activity with the known malicious IP addresses, investigators identified the true entry point: administrative SSO logins to the FortiGate appliance with valid accounts. Another important detail was that there was no evidence of brute-forcing activity for these local accounts. The initial access was established approximately two weeks before any subsequent malicious activity, indicating the attacker used this time to secure consistent access to the environment via the FortiGate device.

Actions such as changing configurations, creating accounts, and downloading configurations might seem harmless individually. However, when viewed together, these activities established a clear pattern consistent with the exploitation of CVE-2025-59718 that facilitated authentication bypass.

Once this groundwork was established through persistence mechanisms and discovery, attackers began authenticating into the environment with their newly created accounts via the SSL VPN connections that led us to investigate the FortiGate device in the first place. These sessions effectively transformed the firewall into an ingress point into the internal network, allowing attackers to move beyond the edge device.

This investigation highlights a common reality in incident response where the first indicator of suspicious activity is rarely the beginning of the story. Instead, responders are often working from a point somewhere in the middle, tasked with reconstructing attacker behavior and peeling back layers of activity to uncover how access was first obtained. 

By following the digital breadcrumbs left behind within available evidence sources, investigators were able to trace the intrusion back to its origin. This process emphasizes the importance of working backward through artifacts and telemetry, recognizing that each piece of data may lead to an earlier stage of attacker activity.

Network edge devices such as firewalls and VPN appliances are often the main vectors of initial access. Despite being critical infrastructure in modern environments, full visibility is rarely achieved in comparison to monitored endpoints. These edge devices can provide valuable evidence during investigations and reveal how initial access went unnoticed.

Conclusion: Key takeaways for defenders

The human element of investigation is crucial. Effective investigations demand a mindset of curiosity; on one side the willingness to dig deeper, and on the other, the ability to look at the big picture. At face value these can seem contradictory, but each facilitates a specific role within an incident response investigation.

Curiosity is what drives responders to grapple with the initial evidence, question assumptions, and identify which threads are worth pulling. It allows responders to move beyond surface-level observations and begin forming hypotheses about what may have occurred. The willingness to dive deeper is what turns those hypotheses into answers. Rather than stopping at the first suspicious event, responders must continue pivoting across logs, correlating activity, and tracing actions further back in time. At the same time, maintaining a big-picture perspective is critical. Individual artifacts or events may appear benign in isolation but when viewed chronologically the attacker behavior emerges.

Looking past any specific incident response methodology, visibility into the environment is essential. Even the strongest investigative approach is limited without access to the right telemetry, thus preventing responders from fully reconstructing an intrusion. In particular, as seen within this investigation, visibility into edge device activity can play a crucial role in unraveling IAV. The network edge is a hostile environment yet is frequently less monitored.

As is often the case with externally facing services and devices, the network edge is constantly targeted. Due to the sheer volume of persistent targeting, this environment can prove difficult to monitor for successful malicious intrusions. Implementing centralized syslog monitoring across these edge devices can close these visibility gaps. It can provide a real-time audit trail of connection attempts, configuration changes, and potential exploit signatures that occur before a threat reaches the internal network.

By effectively pulling on each investigative thread and ensuring visibility across both internal systems and edge devices, defenders can uncover compromises that might otherwise remain hidden. Often, the path to the beginning of the intrusion is already present; it simply requires knowing where, and how, to look.

Detection coverage for Rapid7 customers

Rapid7 actively monitors for emerging threats and leverages evidence from incident response engagements to develop new detection capabilities. Detections have been created and implemented by Rapid7 to pinpoint both exploitation attempts and post-exploitation activities related to FortiGate CVE-2025-59718. For InsightIDR and MDR customers, these detections alert on attacker activity consistent with the techniques described in this blog, enabling earlier identification and response before an intrusion can escalate further.

Detections:

• Potential Exploitation - FortiGate Admin SSO Login and Config Download via External IP

• Exfiltration - FortiGate Config Downloaded Using GUI via External IP

• Suspicious Authentication - FortiGate SSO Login via External IP

Mitigation guidance

Please refer to our initial blog from December, 2025.

MITRE ATT&CK Techniques

Tactic	Technique	Details
Initial Access	Exploit Public-Facing Application (T1190)	Exploitation of vulnerability CVE-2025-59718 on FortiGate firewalls.
Persistence	Create Account (T1136)	Creation of local accounts on FortiGate firewalls.
Persistence and Initial Access	Valid Accounts (T1078)	Use of created accounts and compromised accounts for SSL VPN and RDP authentication.
Defense Evasion	Impair Defenses (T1562)	Firewall rules added to allow for attacker access.
Credential Access	OS Credential Dumping (T1003)	Execution of Mimikatz targeting the local system and Windows Registry hives containing credentials.
Discovery	System Network Configuration Discovery (T1016)	Download of FortiGate firewall configuration files containing sensitive networking information.
Discovery	Network Service Scanning (T1046)	Execution of network scanning tools such as Advanced_Port_Scanner to scan internal IP addresses over SMB protocol.
Lateral Movement	Remote Services (T1021)	Use of Remote Desktop Protocol (RDP).
Execution	Service Execution (T1569.002)	Remote execution of the sysinternals tool PsExec to test credentials against an impacted system.

Indicators of compromise (IOCs)

IOC	Description
Advanced_IP_Scanner_2.5.4594.1.exe	Advanced IP Scanner tool utilized by the attacker.
advanced_ip_scanner.exe	Advanced IP Scanner tool utilized by the attacker.
mimikatz.exe	An open-source post-exploitation tool utilized by the attacker to extract sensitive authentication credentials.
Advanced_port_scanner_2.5.3869.exe	An open-source network utility utilized by the attacker to quickly map active devices and identify open ports.
23.163.8[.]21	Attacker IP address that targeted FortiGate device.
45.32.216[.]250	IP address used by the attacker during FortiGate configuration changes.
45.84.107[.]17	IP address identified in malicious interaction with SSLVPN.
45.80.186[.]84	IP address identified in malicious interaction with SSLVPN.
185.219.157[.]127	IP address identified in malicious interaction with SSLVPN.
185.175.59[.]238	IP address identified in malicious interaction with SSLVPN.
198.98.54[.]209	Attacker IP address that targeted FortiGate device and SSO login.
45.80.184[.]229	Attacker IP address that targeted FortiGate device and SSLVPN.
45.80.184[.]241	Attacker IP address that targeted FortiGate device and SSLVPN.
42.200.230[.]178	Attacker IP address that targeted FortiGate device and SSLVPN.
103.20.235[.]155	IP address identified in malicious authentications to SSO login.
104.28.227[.]105	IP address identified in attacker download of FortiGate configuration file.

What happens when you ditch the tiered ticket queues and replace them with collaboration, agility, and real-time response? In this interview, Hayden Covington takes us behind the scenes of the BHIS Security Operations Center, which is where analysts don’t escalate tickets, they solve them.

The post Inside the BHIS SOC: A Conversation with Hayden Covington  appeared first on Black Hills Information Security, Inc..

But what if we need to wrangle Windows Event Logs for more than one system? In part 2, we’ll wrangle EVTX logs at scale by incorporating Hayabusa and SOF-ELK into my rapid endpoint investigation workflow (“REIW”)! 

The post Wrangling Windows Event Logs with Hayabusa & SOF-ELK (Part 2) appeared first on Black Hills Information Security, Inc..

In part 1 of this post, we’ll discuss how Hayabusa and “Security Operations and Forensics ELK” (SOF-ELK) can help us wrangle EVTX files (Windows Event Log files) for maximum effect during a Windows endpoint investigation!

The post Wrangling Windows Event Logs with Hayabusa & SOF-ELK (Part 1) appeared first on Black Hills Information Security, Inc..

Remember the good ‘ol days of Zip drives, Winamp, the advent of “Office 365,” and copy machines that didn’t understand email authentication? Okay, maybe they weren’t so good! For a […]

The post Stop Spoofing Yourself! Disabling M365 Direct Send appeared first on Black Hills Information Security, Inc..

In this video, John Strand discusses the complexities and challenges of penetration testing, emphasizing that it goes beyond just finding and exploiting vulnerabilities.

The post 5 Things We Are Going to Continue to Ignore in 2025 appeared first on Black Hills Information Security, Inc..

Changes to the msds-KeyCredentialLink attribute are not audited/logged with standard audit configurations. This required serious investigations and a partner firm in infosec provided us the answer: TrustedSec.  So, credit where […]

The post Enable Auditing of Changes to msDS-KeyCredentialLink  appeared first on Black Hills Information Security, Inc..

Recently in the SOC, we were notified by a partner that they had a potential business email compromise, or BEC. We commonly catch these by identifying suspicious email forwarding rules, […]

The post Monitoring High Risk Azure Logins  appeared first on Black Hills Information Security, Inc..

While social engineering attacks such as phishing are a great way to gain a foothold in a target environment, direct attacks against externally exploitable services are continuing to make headlines. […]

The post In Through the Front Door – Protecting Your Perimeter   appeared first on Black Hills Information Security, Inc..

Be sure to read PART 1! Metadata and a New-Fashioned Bank Robbery Let’s face it, some cases are just more interesting than others and, when you do incident response for […]

The post OSINT for Incident Response (Part 2) appeared first on Black Hills Information Security, Inc..