Cybersecurity Researchers Slam Guardrails on Anthropic's Fable

Anthropic's latest model Fable has been met with criticism from cybersecurity researchers who claim its guardrails are too restrictive, limiting the model's capabilities and creating frustration among users.

Fable Rejects Innocuous Requests

Fable has been designed to reject any request that could be tangentially related to cybersecurity or biology. This includes innocuous tasks like reading a blog post. According to Valentina "Chompie" Palmiotti, a well-known security researcher at IBM X-Force, Fable pauses the chat and says that its "safety measures flagged this message for cybersecurity or biology topics."

This restriction has led to frustration among users, with some complaining that even asking for a code review triggers Fable's guardrails.

• Matt Suiche, a cybersecurity veteran, told TechCrunch that "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."

• Suiche also noted that Fable's guardrails seem to be keyword-based, triggering the model to fall back to Claude Opus 4.8 if it hits a guardrail.

Anthropic's guardrails were put in place to limit the risk that Fable could be used to develop malware or compromise software. However, many cybersecurity experts are still put off by the haphazard nature of the restrictions. Some have called for Anthropic to relax the guardrails over time, allowing users more flexibility and capability.

Cybersecurity Professionals Must Apply to Cyber Verification Program

Apart from guardrails inside its models, Anthropic requires cybersecurity professionals to apply to the Cyber Verification Program. If they get approved, the applicants have fewer limitations on using Claude for cybersecurity work. OpenAI has a similar program called Trusted Access for Cyber.

Despite the criticism, Anthropic has maintained that its guardrails are necessary to ensure the safe and responsible use of its models. However, the backlash from cybersecurity researchers highlights the need for more nuanced and transparent guardrails that balance safety with capability and flexibility.