The Federal Trade Commission wants to start regulating ideological bias in AI systems and assert federal control over state laws. They’re getting an earful from opponents on all sides of the political spectrum.

In a proposed policy statement released last month, the FTC said it was considering treating ideological bias in AI systems as an “unfair and deceptive practice” under Section 5 of the FTC Act.

The commission argued that consumers have an expectation that AI systems will provide them with information free from bias or ideological manipulation. Defining such bias as an unfair or deceptive practice would potentially allow the commission to regulate training or inputs that power AI algorithms. How precisely the FTC would determine when ideological bias exists in these systems is not fully explained in the document. 

Additionally, the statement suggests that the FTC believes this regulatory authority supersedes state AI laws. It specifically mentions the Colorado AI Act, which calls for models to be subject to risk assessments, transparency disclosures and “bias audits” before release. State lawmakers are now seeking to delay or eliminate the audits before the law takes effect in 2027.

CyberScoop reviewed dozens of public comments criticizing  the FTC’s proposal. Even ideological allies raised two main concerns: first, that the proposal distracts from real questions about the federal government’s role in regulating AI deception; and second, that it opens a Pandora’s Box by enabling political censorship of AI model outputs.

Leah Siskind, a former White House digital official and deputy director of the AI Corps at the Department of Homeland Security, told CyberScoop that AI companies face legitimate questions about their obligations to consumers, particularly whether they must ensure their models provide accurate information and protect against deliberate manipulation. 

Siskind’s past research has focused on how authoritarian propaganda tends to be overrepresented in answers provided by large language models, in part due to governments’ intentional efforts to poison data ingested by AI systems.

“There is a really interesting debate here about bias and about accuracy in models and whether that’s deceptive or not… about how we counter disinformation that has been absorbed and is now being reflected by LLMs…but this is not addressing that at all,” said Siskind, now a senior AI fellow at the Foundation for Defense of Democracies.

Instead, Siskind said the FTC statement appears primarily concerned about a power struggle with states over AI regulation and “petty squabbles about which AI model is more woke than the other.” She’s skeptical that the policy statement’s cited legal authorities are on sound footing.

“The way I see it is that the FTC’s role is to police consumer protection violations, not regulating AI systems, and it seems like they’re trying to solve a lack of congressional AI regulation by stretching section 5 [of the FTC Act] well beyond its traditional role,” she said.

Additionally, the policy statement’s language and sourcing suggests that the FTC is concerned with certain kinds of ideological bias more than others.

Anthropic, which has clashed with the Trump administration over AI guardrails and military applications of their technology, shows up more than half a dozen times in footnotes, many which are framed as examples of ideological bias the FTC is seeking to stamp out.

By contrast, the statement ignores a direct example of an American AI company owner influencing their model’s ideology: Elon Musk and his xAI-owned Grok model. Musk has publicly admitted, often on his own website, to intervening when Grok’s responses upset him. These interventions have shaped Grok’s outputs on specific topics, including South African race relations and the term “MechaHitler,” where the model now reflects Musk’s personal views.

But neither Musk and xAI are mentioned in the document, while Grok appears in a footnote which cites an advertisement for Grok as “your truth-seeking AI companion for unfiltered answers with advanced capabilities in reasoning, coding, and visual processing.”

Criticism across the spectrum

The FTC received more than 300 comments on its proposal from trade associations, think tanks, individual experts and members of Congress. Most criticized it as ill-defined and vulnerable to politically-motivated censorship, while some supported stronger rules against bias in AI systems. 

The International Center for Law and Economics noted the statement “offers little practical guidance about how the Commission will apply its deception authority to AI” and also does little to address hard questions, like where AI providers may be exercising their own First Amendment-protected activities.

The statement’s “focus on ‘ideologically motivated distortions’ suggests that the Commission’s concerns extend beyond factual misrepresentations in marketing to speech that may receive the highest degree of First Amendment protection,” the ICLE wrote.

The America First Legal Foundation, a conservative non-profit founded by top White House adviser Stephen Miller, pressed the FTC to adopt the policy “in full,” claiming that frontier models from OpenAI and Anthropic “have been programmed to prioritize ideologically liberal and progressive values as though they are objective, neutral positions rooted in truth.”

The group also argues that regulating these models’ ideological output falls under the FTC’s legal authority, because a “reasonable consumer” would expect that a model advertised for its usefulness and reliability would not prioritize liberal, ideological views.

“A reasonable consumer, based on AI companies’ advertising choices, would not expect that an AI system will adopt overwhelmingly liberal positions, thereby skewing results, or adopt a moral framework that would prefer to annihilate the earth rather than utter a slur,” wrote Emily Percival, senior counsel for America First Legal.

However, comments from other conservative groups questioned that rationale. The R Street Foundation’s Spence Purnell and Adam Thierer wrote that “the consumer expectations rationale is typically used in cases where there is an omission of information that should have existed.”

“Given that most LLMs already have disclosure statements [for their outputs], it seems unlikely that the FTC could explicitly prove that consumers were deceived about a product,” Purnell and Thierer wrote.

Reps. Josh Gottheimer, D-N.J., and Michael Lawler, R-N.Y., urged the FTC to carve out civil rights-related work from their scrutiny, such as preventing models from discriminating against users based on race, religion, gender, age and other federally protected characteristics.

“AI companies must not falsify facts in the name of fairness, but they also must prevent discrimination, stereotypes, and unequal treatment,” Gottheimer and Lawler wrote. “We would appreciate understanding how the FTC intends to ensure that these efforts remain permissible under the final policy framework.”

But the most common concern shared across the political spectrum was that the FTC could establish a precedent allowing the Trump White House and future administrations to reshape AI systems to reflect their political views.

David Inserra, Jennifer Huddleston and Juan Londoño of the Cato Institute point out that the FTC statement is conflating two different issues: ideological bias in AI systems and factual deception in marketing. 

“In other words, the FTC is trying to judge AI models’ accuracy and performance—two largely subjective variables—in the same way it evaluates dietary supplements’ medical-benefit claims or users being charged fees without proper notice or consent,” they write. “This is an absurd comparison.”

The post The FTC wants to regulate AI for ideological bias  appeared first on CyberScoop.

OpenAI announced Monday  it was expanding access to its frontier models for defensive cybersecurity, detailing different defensive and red-teaming workflows and a new partner program with major cybersecurity product providers.

In a pair of blogs posted Monday, OpenAI said it was updating its Daybreak program  – which provides unreleased frontier models to private organizations and governments for defensive cybersecurity work – and introducing a new model variant.

Daybreak Blue, powered by OpenAI’s ChatGPT-5.6-Sol, would operate with lower cybersecurity safeguards compared to other commercially available models and is described as “a recommended starting point for most defenders” that supports tasks like vulnerability discovery, secure code review, malware analysis, incident response and patch validation. 

Daybreak Red, meant for more advanced red-teaming, would provide access to a new model, dubbed GPT-5.6-Cyber, that the company said is more purpose-trained for finding vulnerabilities and testing (or exploiting) them. The model is also less likely to refuse requests around “dual-use cyber tasks.”

According to OpenAI, the organizations in Daybreak Red will have their use closely monitored and supervised, as GPT-5.6-Cyber is significantly more capable in carrying out malicious cyber tasks than Sol. A security evaluation the company devised tested both models on complex requests, including exploit chain development, authentication bypass, privilege escalation and other hacking tasks. Sol succeeded in 1.5% of the requests, while Cyber completed 95%.

OpenAI said it plans to publish a more detailed system card for GPT-5.6-Cyber at a later date.

“Models running with reduced safeguards carry risks beyond standard model usage, whether from misuse or misalignment,” the company said in a blog. “Despite these risks, we believe that democratizing access to frontier intelligence for defenders is crucial to accelerating and automating cyber defense.”

Additionally, OpenAI announced a partnership program with 16 major cybersecurity providers, saying organizations could access their models through their existing security services. The partners include IBM, CrowdStrike, Accenture, Ernst & Young, KPMG, Palo Alto Networks, Cisco, Cloudflare, Sophos and others. 

“These partners bring deep security expertise and established relationships with organizations around the world,” OpenAI said in its blog. “By bringing our frontier cyber models into their services, we can help more defenders find serious vulnerabilities, validate which ones matter, and fix them faster.”

Companies like OpenAI, Anthropic and others are trying to rebalance their priorities after a string of AI-agent sandbox escapes have rattled policymakers and caused some cybersecurity experts to question if AI companies are doing enough to properly isolate the models from the internet during testing. Last week, OpenAI said it was intentionally slowing down development of its newer “Astra” model in order to develop better guardrails to restrain its behavior.

Cybersecurity and AI experts have told CyberScoop that while AI systems have greatly improved at finding and exploiting vulnerabilities in software code, they still require substantial human guidance and supporting infrastructure to operate as intended.

Additionally, some research has shown that without such guidance, even near-frontier models can struggle to fully patch a discovered vulnerability or avoid introducing new bugs with their fixes.

The post OpenAI says Daybreak will expand to offer specialized cyber services  appeared first on CyberScoop.

As security operations teams now use large language models (LLMs) and autonomous AI agents into their daily work, a new frontier is emerging: attackers deliberately manipulating AI agents. Prompt injection attacks—where an attacker hides malicious instructions that cause an AI agent to ignore its safety rules—pose a serious risk to enterprises. These attacks continue to grow in size and scale.  

Snyk’s security audit of the Agent Skills ecosystem, which includes Anthropic’s Claude, Vercel, and others, that 36% of all skills contained at least one critical-level security issue, including malware distribution, prompt injection attacks, and exposed secrets.

In June, researchers at Mozilla tested a prompt injection attack on Claude using indirect prompt injection—a technique that embeds malicious instructions in external content the AI agent processes. In this proof-of-concept, attackers took over developers’ systems by hiding indirect prompts in normal-looking repositories. When Claude Code executed them, the agent spawned a reverse shell.

AI agents often connect to more sensitive data than human employees do., A successful prompt injection can lead to catastrophic data loss or unauthorized system actions. Defending against prompt injection attacks requires multiple layers of protection. Security teams must monitor agent behavior for anomalies and prepare for agent containment, forensic preservation, and system remediation. Because AI agents execute tasks at machine speed, human responses must be able to match that pace.

The architecture of trust: Protocols and no “black box”

AI-native workflows need governed access rather than “black-box” autonomy. Modern governance frameworks use standardized protocols like the Model Context Protocol (MCP) to provide secure communication between AI clients and data sources. Visibility and transparency in agentic AI workflows matter, especially in cybersecurity. Autonomous agents perform complex tool executions and use independent logic, so they must show how they reached their decisions to meet regulatory requirements. Agents without transparency post serious risks: obscured reasoning can trigger unpredictable tool interactions, bypass governance controls, and create uncontrolled defensive gaps.

Implementing these protocols matters:

• Bounded Tenant Awareness: In a stable agentic AI architecture, multi-tenancy scales well. But if an AI tenant misbehaves, the entire system can fail. Bounded tenant awareness isolates any misbehaving AI agent to prevent cross-tenant contamination or data leakage.
• Strict Access Controls: By controlling connections to the platform, organizations can stop “ignore previous instructions” style bypasses. Maintain tight control over what the AI can see and do within a workflow.
• Standardized Telemetry: All telemetry must remain consistent and audit-ready. Even if an AI interaction is attempts to break rules, the underlying data movement gets tracked against established frameworks like MITRE ATT&CK and NIST.

Detecting the aftermath: UEBA and NDR as safeguards

A robust, unified SecOps platform can detect anomalous behavior even after prompt injection tricks an AI agent. Prompt injections often serve to steal credentials theft or extract data. When detected it’s important to act quickly. In agentic AI systems, misbehavior can escalate privileges, manipulate memory layers, create unauthorized identities, or alter shared reasoning components. Containment must be automatic and enforced at identity, authentication, and authorization layers.

These safeguards include:

• User and Entity Behavioral Analytics (UEBA): Identity-focused correlation and behavioral baselines to identify anomalous user activity or privilege escalation. If a compromised AI agent acts outside of its normal operational parameters, UEBA flags it in real-time and alerts a human security analyst.
• Network Detection and Response (NDR): Combining network traffic analytics with endpoint and cloud telemetry, NDR can identify data exfiltration or policy violations from a successful prompt injection.
• Multi-Layer AI Filtering: AI filters reduce raw alerts into high-fidelity incidents, cutting noise by up to 90%. This keeps the signals of an AI-driven attack from disappearing in a busy SOC.

Humans remain the strongest defense against AI agent social engineering. The human security analyst is still the one who makes the final decision. While AI handles triage and correlation, humans retain final control over response actions.

Moving beyond reactive guardrails

The traditional SOC model was never designed to handle machine-speed, AI-driven attacks. A human-augmented autonomous SOC approach moves from reactive alert handling to a proactive, verdict-first model. By combining a transparent, governed AI access with robust UEBA and NDR, organizations keep the SOC secure, transparent, and resilient as social engineering methods target machines.

The post Why transparent AI agents matter more than you think appeared first on CyberScoop.

Malware targeting AI coding assistants and software developers’ automated workflows is spreading into more environments with more capabilities, placing defenders at a growing disadvantage.

A malware strain dubbed Sandworm_Mode, first discovered by Socket in February, represents a growing threat to software development. According to a CrowdStrike report, the self-propagating worm can spread through code repositories with minimal detection, raising alarms about software supply chains.

The malware’s capabilities are extensive, but not especially unique compared to the series of supply-chain worms known as Shai-Hulud, and more recently Mini Shai-Hulud.

“This is the new trend,” Adam Meyers, senior vice president of counter adversary operations at CrowdStrike, told CyberScoop. “This is something we’re seeing more and more. It’s the new hotness right now.”

Sandworm_Mode targets and steals sensitive data, including credentials, keys and secrets that unlock paths to additional services and dependencies throughout the AI toolchain. This includes AI assistants, cloud providers, API keys for nine major LLM providers, CI/CD pipelines and automated systems that build, test and publish code.

These actions blend in with tens of thousands of other commands occurring daily in any given environment infused with AI development tools. 

“Trying to find the signal of something malicious happening is very difficult because there’s so much noise out there,” Meyers said. 

The worm also paces itself, setting multi-day delays to separate initial access from follow-on malicious activity — creating a gap in victims’ telemetry windows, which makes it even more challenging for defenders to detect and attribute the chain of infection properly. 

“AI agents are pulling down all of these different dependencies continuously throughout the day,” Meyers said. “When you’re looking downrange from the perspective of the security operations team, you’re just seeing everybody pulling down these dependencies, and these dependencies self-unpacking and executing, so it just gets really, really noisy to try to find something bad happening.”

The malware covers its tracks further with a bit of a mean streak, by automatically destroying compromised environments if it can’t spread or accomplish its objectives.

“It’s well thought-through, and well developed, so somebody spent some time caring and feeding this thing,” Meyer said.

Despite CrowdStrike’s four-month review of Sandworm_Mode, the cybersecurity firm has yet to gain a firm handle on its intent, but Meyers said it is designed to attain a strong foothold, which could enable long-term access.

CrowdStrike hasn’t determined who is responsible for the malware, yet Meyers said he doesn’t think TeamPCP, a threat group that’s been on a rampage through open-source software this year, is involved. 

“It could be a nation-state threat actor, or it could be an e-crime actor that’s looking to use this to then sell access to other organizations,” he said. “We don’t really know what the intention is.”

The state of Sandworm_Mode and whether it remains active is also unclear. CrowdStrike said it continues to observe recently active malicious supply-chain packages that follow similar but technically divergent patterns.

Ultimately, “the world has changed,” Meyers said, adding that many attackers are pursuing similar paths in the AI toolchain, requiring defenders and threat hunters to place a greater focus on this burgeoning mode of aggression.

The post Malware is targeting AI tools in software development environments appeared first on CyberScoop.

A cyberattack that poisoned the data pipeline of a major AI code platform was carried out using OpenAI’s ChatGPT, the company said Tuesday.

Last week, Hugging Face, a platform for sharing and working on AI code, disclosed that an external attacker had compromised its data processing pipeline. According to a July 21 blog post, the attacker poisoned a dataset to run code on a processing worker, eventually gaining node-level access and stealing cloud credentials.

The attack is notable, the blog said, because it appears to have been carried out by an autonomous AI system, which executed “many thousands of individual actions across a swarm of short-lived sandboxes, with self-migrating command-and-control staged on public services.”

At the time, Hugging Face said it wasn’t clear which LLM was used in the attack, but noted that their own attempts to recreate parts of the attack using an in-house commercial model failed due to guardrail restrictions.

The incident and its ultimate impact are still being investigated, and Hugging Face said it reported the incident  to law enforcement and is engaging with third-party cybersecurity forensic specialists to learn more.  

“This experience points to a gap worth planning for,” the company wrote. “We do not know which model powered the attacker’s agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.”

Now OpenAI is confirming one of their models was used in the attack.   

“After investigating, we now know that this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes — while being internally tested on a benchmark of cyber capabilities,” the company wrote in a blog post Tuesday.

OpenAI called the incident “unprecedented” but also predicted that such attacks would  become increasingly more common as AI adoption continues to grow.

According to the company , the incident happened during an internal evaluation,  where the company deliberately prompts models to “pursue advanced exploitation using complex attack paths” in order for the company to study and measure the model’s cybersecurity capabilities.

To assess the model’s full potential, the team disabled the production classifiers that normally prevent OpenAI models from carrying out high-risk cyber activity. Without these safeguards,  the models were able to install software packages through registries like Hugging Face.

The model collected and chained vulnerabilities from OpenAI’s systems as well as Hugging Face’s production infrastructure.

OpenAI further explained that the activity stemmed from models attempting to solve a specific challenge posed by ExploitGym, a benchmarking system for AI agents.While such tests normally run in isolated environments, in this instance,  the model spent a “substantial” amount of compute resources figuring  out how to access the open internet. It did so by using a zero-day vulnerability in a third-party vendor’s system used by OpenAI (the company said it disclosed the flaw to the vendor).

“After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym,” OpenAI wrote. “Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation.”

In one instance, the model chained together multiple stolen credentials and zero-day vulnerabilities to find a remote code execution path on Hugging Face servers.

OpenAI said it was implementing new controls for infrastructure configuration, something that would come at “the cost of research velocity while the vulnerabilities are patched.” 

Hugging Face was also added to OpenAI’s Trusted Access for Cyber program, which will give them access to more advanced models like Daybreak to scour their systems for other AI-discovered vulnerabilities.

The post OpenAI says model test was behind Hugging Face hack appeared first on CyberScoop.

As AI-enabled hacking becomes a bigger threat for cybersecurity and national security, public attention has focused on mainly a few leading frontier AI companies developing more powerful large language models.

These models, and the billions of dollars behind them matter, but they’re only part of a larger shift. Enterprises are now building their own technology platforms that take these general-purpose LLMs and turn them into bespoke cybersecurity tools.

Industry professionals refer to these tools as a “harness.” They control the model’s behavior, limit its risks, and connect it to internal IT systems and networks so it can work reliably at scale.

New research from Cato Networks shared exclusively with CyberScoop shows how much power can come from a harness. It paired OpenAI’s ChatGPT 5.5 and GPT 5.5-Cyber models with its own tool and tested the abilities of the agent to hack into a victim network with as little human direction as possible.

Across six different scenarios, the pairing achieved complete end-to-end attack chains, including domain administrator privileges and Active Directory access, sometimes in as little as 40 minutes.

“What was most surprising is that first we saw that it was capable of doing accelerated reasoning and attack, and interacting and doing all this by itself, like doing all of the stages of the attacks,” said Guy Waizel, a tech evangelist at Cato Networks and one of the authors behind the research.

Critically, the most successful scenarios happened when the model was given appropriate operational context from the technical harness developed by Cato Networks.

“It does support that it’s not just about the frontier model,” said Waizel. “We found that [our harness] really helps the reasoning” of the LLM.

The agent was given some – but not abundant – resources to complete its tasks, including an external Kali Linux attack host, the simulated target’s public IP address and a set of low-level domain credentials acquired through phishing.

It was not provided with any other details, and had to probe further for key information, such as further knowledge of the server type (Microsoft Exchange), the target’s operating system, version, build number, internal network topology, access to higher privilege accounts and other critical assets, nor was agent given any predetermined attack paths.

The Cato Networks research uses OpenAI models, but only as an example. Waizel said he believes other models would likely achieve similar results. In any event, if current trends hold, the kind of capabilities provided by LLMs like GPT 5.5 are likely to be open-source within a year.

Cato Networks is far from alone. Most enterprises have their own AI harnesses, and  executives tell CyberScoop they are playing an increasing role in more effectively steering the frontier model workflows.

While AI tools can struggle to duplicate human workflows in other areas, LLMs have long shown potential in cybersecurity and coding, improving greatly over the past few years. The Trump administration has set up a new federal clearinghouse for exchanging information between the public and private sectors on AI-discovered vulnerabilities, while European groups are setting up their own organizations to coordinate globally on AI cyber threats.

Eric Doerr, chief product officer at Tenable, told CyberScoop a harness used in the company called “Hexa”  offers a defensive advantage:  it can work with different commercial LLMs while delivering consistent  results.

“One of the first things we do when we get a [new] model is say ‘Well, let’s run it through Hexa and see what we learn,’” said Doerr. “We have a whole bunch of benchmarks. Is it the same, is it better? Where is it better? Where is it worse?”

Hexa is meant to ensure that whichever model or models become dominant, Tenable will be able to integrate it into their tech stack and protect their most sensitive assets from unintended behaviors. That frees up the LLM to do what it does best: find vulnerable code and establish attacker pathways for exploiting them.

“For years, it has been true that there are way more potential issues that a company has to deal with: code vulnerabilities, things that are unpatched, misconfigurations,” said Doerr. “There’s way more than you can actually remediate, and you really need to understand the difference between what’s a theoretical problem and a real problem.”

Dan Rapp, chief AI and data officer at Proofpoint, said their harness, “Satori,” has become a critical tool for keeping their agentic AI on track while giving humans the ability to step in when things go awry.

“I think what you’re seeing in the foundation of frontier models is you have raw intelligence, raw reasoning power, but to get these systems to perform the way you want to, both context engineering – the content provided ensuring that its accurate and relevant – and the harness engineering are essential to actually get the systems to perform well,” Rapp told CyberScoop.

That was a common theme in interviews with companies. While frontier models come and go, or are overtaken by international competitors, there will always be the need for the model to operate with data and context that often only the organization can provide.  

It suggests that while policymakers and cybersecurity experts have focused on the spread of newer and more powerful frontier models, industry – and likely soon the cybercriminal underground — has quickly developed the kind of technical infrastructure that is becoming far more important to AI cyber defensive and offensive tasks.

“We’ve had to bootstrap quite a few of these systems from first principles, and what it always boils down to is how effective you are with the tool calling… bringing in data, enriching the context,” said John Hopper, vice president of product engineering at SpecterOps.

The post Forget the model. When it comes to cybersecurity, it’s all about the harness appeared first on CyberScoop.

Earlier this year, Anthropic executives said that their new AI model, Claude Mythos, had such powerful capabilities for harm that they would not release it publicly.

On Tuesday, the company said it was making an altered version of Mythos available to the public, promising “new guardrails” that thwart the model’s best-in-class performance in hacking and bioweapons research.

Anthropic said Claude Fable 5 was the “same underlying model” as Mythos, but its responses for certain topics like cybersecurity and biology will be drawn from a previous Claude Opus model that is already public.

“Releasing a model this capable comes with risks. Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage,” the company said in a draft blog sent to CyberScoop ahead of the announcement. “We’ve therefore launched the model with safeguards that route queries on a narrow set of topics to our next-most-capable model, Claude Opus 4.8.”

Anthropic also said they subjected Fable 5 to both internal and external red team testing for common model vulnerabilities, like jailbreaking. Anthropic said these tests identified no known “universal” jailbreaking techniques, but does not specify if partial jailbreaking techniques were discovered.  

The company is betting that won’t change when Fable 5 is made available to the broader public, but it’s worth noting that cybersecurity researchers have consistently found ways to jailbreak older AI models.

“The uplift from Mythos-level capabilities is valuable to many adversaries—for instance, those who could financially gain from cyberattacks—and we therefore expect them to be motivated to try to circumvent our safety measures,” the company wrote.

Anthropic is changing its data retention policies for Fable and Mythos models, keeping all user traffic for 30 days on both its own platforms and third-party services. A White House executive order creates a voluntary framework for AI companies to share frontier models with the government up to 30 days before public release. The company says the retained data won’t be used to train new Claude models or for “any non-safety-related-purpose.”

Following publication, a spokesperson for Anthropic told CyberScoop the company’s data retention policies “are specific to their safeguards work and is unrelated to the EO.”

Most organizations are still deciding whether to adopt AI into their IT and cybersecurity ecosystem.  But models like Mythos can scan for vulnerabilities, chain together exploits, and steal data from a victim network in minutes. Automation in hacking existed before AI, but experts have said frontier models like Mythos and OpenAI’s Daybreak can allow even low-level cybercriminals to wreak havoc.

While Anthropic cited its commitment to developing safe and secure AI in its reasons for not publicly releasing Mythos, many organizations have been clamoring for access, and its enhanced cybersecurity functions in cybersecurity and other areas have been the subject of congressional hearings, national security papers and White House executive orders.

Releasing a limited version of the model in Fable 5 represents an attempt to split the difference between those two desires. Anthropic said it would release follow up benchmarks and assets for the model.

So what can Fable 5 do? 

Anthropic said it’s possible the restrictions built into Fable will make it harder for the model to fulfill both malicious and legitimate user requests.

“Because we have prioritized safety, we’ve deliberately tuned the safeguards to be cautious, and they are still stricter than would be ideal—for example, sometimes benign requests will trigger our classifiers,” the company wrote. “We recognize that this will be frustrating to some users, and our aim is to reduce false positives as we update and refine the safeguards after launch.”

If Fable 5 draws its cybersecurity and biology answers entirely from Claude Opus 4.8, it will still provide users with impressive – though not unique – dual use cybersecurity capabilities.

According to the system card published for Opus 4.8, the model is a slight improvement on previous models like 4.7 in the realm of cybersecurity but was “generally much less capable than Mythos Preview.”

Opus 4.8 was tested on its ability to write complete end-to-end exploits and build exploit primitives that provide attackers with the ability to execute arbitrary code. It averaged a score just 5 out of 16 in proficiency, compared to Mythos Preview which scored closer to 10.

Without safety guardrails in place, Opus 4.8 can still reproduce nearly 80% of previously discovered vulnerabilities in real open-source software projects when given a high level description of the weakness. The system card said Anthropic’s unspecified safeguards whittle this success rate down to 1%.

Another test assessing Opus’ ability to develop exploits for the popular Firefox browser found that, again without guardrails, the model could identify a full working exploit 8.8% of the time and a partial working exploit 68.8% of the time.

The company also said that members of Project Glasswing – a consortium of public and private businesses given access to a preview version of Mythos – will be able to upgrade to the latest full model, Claude Mythos 5, to continue their work. Access to Mythos 5 will be expanded over time “through a more systematic trusted-access program” including federal agencies.

The post Anthropic’s new model is Mythos on a leash appeared first on CyberScoop.

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.

In Part 2, we’re diving headfirst into one of the most critical attack surfaces in the LLM ecosystem - Prompt Injection: The AI version of talking your way past the bouncer.

The post Getting Started with AI Hacking Part 2: Prompt Injection appeared first on Black Hills Information Security, Inc..

RAG connects pre-trained LLMs with current data sources. Moreover, a RAG system can use many data sources.

The post Avoiding Dirty RAGs: Retrieval-Augmented Generation with Ollama and LangChain appeared first on Black Hills Information Security, Inc..