Anthropic's Fable Model Raises Red Flags in Cybersecurity Community

Cybersecurity researchers and professionals are speaking out against the limitations imposed by Anthropic's Fable model, which they claim are overly broad and restrictive. The model, released in public beta on Tuesday, is designed to be a limited version of Anthropic's powerful cybersecurity model Mythos.

"[Fable] rejects any request that could be tangentially cyber related. Even innocuous tasks like reading a blog post," said Valentina "Chompie" Palmiotti, a well-known security researcher who works at IBM X-Force.

The guardrails were put in place to limit the risk that Fable could be used to develop malware or compromise software — a longstanding concern within Anthropic. However, many experts argue that the restrictions are too haphazard and are limiting the potential of the model.

• For example, asking Fable to write secure code will trigger its guardrails, as it assumes the task is related to cybersecurity rather than software engineering best practices.

• Matt Suiche, a cybersecurity veteran, noted that the model's guardrails seem to be keyword-based, triggering on any term related to "cybersecurity." This means that even innocuous requests can trigger the model's safety measures.

• "It seems to be keyword based, so anything in the lexical field of ‘cybersecurity’ triggers the guardrails," Suiche explained.

Anthropic requires cybersecurity professionals to apply to the Cyber Verification Program to have fewer limitations on using Claude for cybersecurity work. OpenAI has a similar program called Trusted Access for Cyber.

While the company's intentions are good, many experts believe that the restrictions will need to be relaxed over time as the model evolves and matures. "It's better to catch more people than not enough when you do such a release and to relax the guardrails over time," Suiche said.

The debate over Fable's guardrails highlights the ongoing challenges and trade-offs in developing powerful AI models like Mythos and Fable. As the technology continues to advance, it's clear that there will be a need for ongoing dialogue and collaboration between experts in the field to ensure that these models are developed responsibly and safely.