Researchers criticize guardrails on Anthropic’s new Fable model

On Tuesday, Anthropic released Fable, a public and limited version of its cybersecurity model Mythos, which was originally released in April. The launch has faced immediate pushback from cybersecurity professionals regarding its restrictive safety measures. Anthropic implemented guardrails—safety restrictions programmed into an AI model—to limit the risk that Fable could be used to develop malware or compromise software. Additionally, restrictions on biology topics are intended to prevent the development of biological weapons. However, cybersecurity researchers are unhappy with these restrictions, arguing they severely limit the model’s utility.

Security researchers report that Fable’s guardrails are overly sensitive and disrupt legitimate work. Valentina “Chompie” Palmiotti, a security researcher at IBM X-Force, stated that “[Fable] rejects any request that could be tangentially cyber related. Even innocuous tasks like reading a blog post,” which limits its utility for legitimate security work. Matt Suiche, a cybersecurity veteran and member of the technical staff at Tolmo, an AI cybersecurity startup, explained that when users ask the model to write secure code, it assumes the request is cybersecurity-related work rather than standard software engineering best practices, resulting in a downgrade. Suiche added that the guardrails appear to be keyword-based, meaning any terms within the lexical field of cybersecurity trigger the safety measures. Another researcher noted that even requesting a code review triggers the guardrails. When Fable hits a guardrail, it falls back to Claude Opus 4.8, the fallback model used when guardrails are triggered.

Despite the friction, Suiche noted that the strict guardrails are understandable in the early days of the release, suggesting they will evolve as AI companies collaborate more with cybersecurity firms. To bypass these limitations, Anthropic requires cybersecurity professionals to apply to its Cyber Verification Program to obtain access with fewer restrictions. This approach mirrors competitor OpenAI, which operates a similar program called Trusted Access for Cyber. Meanwhile, Anthropic has expanded access to the more powerful underlying Mythos model. After initially restricting Mythos to a limited number of organizations under Project Glasswing—an effort to deploy the model to secure critical software and infrastructure—in April, Anthropic expanded access last week to hundreds of organizations across 15 countries.

Why it matters

The tension between safety guardrails and model utility highlights the ongoing challenge for AI labs in balancing risk mitigation with the practical needs of cybersecurity professionals.