Anthropic unveils Claude Fable 5 but reserves full cyber capabilities for vetted users

Researchers describe model's safety controls as overly restrictive, interfering with legitimate security work

AI company Anthropic has launched its most capable AI system to date, while introducing an unusual set of safeguards designed to prevent its most powerful cybersecurity features from reaching the wider public.

The new model, Claude Fable 5, became generally available on Monday through Anthropic's API and subscription services.

The company has simultaneously released a second version, known as Claude Mythos 5, which it says possesses the same underlying capabilities but is reserved exclusively for vetted cybersecurity professionals and operators of critical infrastructure.

Anthropic claims Mythos 5 is currently the world's most powerful cybersecurity-focused AI model.

One model, two versions

Unlike previous AI launches, the distinction between Fable 5 and Mythos 5 is not based on the intelligence of the models themselves.

Instead, Anthropic has built a layer of monitoring systems around Fable 5 that detect potentially dangerous requests involving cybersecurity, biology, chemistry, and AI model distillation techniques.

When such requests are detected, users are automatically redirected to a less capable model, Claude Opus 4.8, rather than receiving an answer from Fable 5.

Anthropic says users are informed when this handover occurs.

The company argues that unrestricted access to Mythos-level cyber capabilities could significantly increase the effectiveness of malicious actors by automating tasks such as network reconnaissance, vulnerability discovery and exploitation.

Testing the safeguards

Anthropic says the protective system relies on separate AI classifiers trained to identify misuse attempts and jailbreak techniques.

According to the company, the classifiers were intentionally tuned conservatively for the initial release, acknowledging that some harmless requests may be incorrectly flagged.

Anthropic estimates that fewer than 5% of user sessions trigger a fallback to Opus 4.8, meaning that for the vast majority of interactions Fable 5 behaves similarly to the unrestricted Mythos 5.

Frustration among researchers

Anthropic's restrictions have drawn criticism from some cybersecurity researchers who say the system's safety controls are overly restrictive and interfere with legitimate security work.

Several researchers reported that Fable blocks or downgrades responses to requests they describe as routine and harmless.

Valentina "Chompie" Palmiotti, a security researcher at IBM X-Force, wrote on social media that the model "rejects any request that could be tangentially cyber related", including what she characterised as innocuous tasks such as analysing blog posts.

Some cybersecurity professionals say the filtering system appears to be overly broad.

Matt Suiche, a veteran security researcher and member of the technical staff at AI security startup Tolmo, said even requests focused on secure software development could trigger the safeguards.

"If you ask it to write secure code, it assumes it is cybersecurity-related work instead of software engineering best practices, and you get downgraded," he told TechCrunch.

"It seems to be keyword-based, so anything in the lexical field of cybersecurity triggers the guardrails," he said.

Other researchers reported similar experiences online, with some claiming that even requests for code reviews caused the model to activate its restrictions.

Anthropic has indicated that its safeguards may be refined in future updates as it gathers feedback from users and studies how the system performs in real-world conditions.

More industry reaction

Commenting on Fable 5 launch, Jon Abbott, chief executive of cybersecurity firm ThreatAware, said: "It is difficult to tell the real difference and impact of the first Mythos-class model compared to Opus 4.8. Some are saying the difference from a security perspective isn't that great. However, I think this is largely irrelevant as even the current models are quicker than most defenders can handle. We can also be confident that the models will become exponentially better at finding vulnerabilities."

Abbott argued that organisations should remain focused on fundamental security controls, including endpoint detection tools, multi-factor authentication and web filtering.

"If organisations fail on the basics, the difference between minutes and hours in exploit-to-attack timeframes becomes largely irrelevant," he said.

Others questioned whether the significance of Mythos 5 has been overstated.

Chris Wallis, chief executive of security firm Intruder, said independent observers have limited visibility into the model's true capabilities because Anthropic has not made unrestricted versions publicly available.

"The problem is we don't know how closely it actually resembles 'Mythos Preview'," he said. "The whole thing is behind a Wizard of Oz-style curtain."

Wallis also pointed to research suggesting rival systems may already outperform Anthropic's earlier cyber-focused models in some attack simulations.

"What will be more interesting is when AI models without guardrails catch up, open source is generally only a few months behind the major models - and whether that will start to change the game significantly," he said.

Earlier this month, Anthropic announced its plans to go public, with an IPO planned for October, with previous funding round valuing it at around $965 billion.