Anthropic: Experts Criticize Restrictions on AI Models

Strong Opposition to the American Decision

A group of 76 cybersecurity experts, including prominent figures from the industry, has sent an open letter to the U.S. government. They are calling for the lifting of export restrictions imposed on the Fable and Mythos AI models developed by Anthropic. According to these experts, these models are crucial for detecting vulnerabilities and securing software.

The letter emphasizes that depriving cybersecurity defenders of these tools, while threats are evolving rapidly, is a dangerous decision. Last Friday, the government ordered Anthropic to limit the export of these models, citing national security concerns without providing details. As a result, Anthropic suspended access to Fable and Mythos for all users worldwide.

Renowned Experts Speak Out

Among the signatories of the letter are notable figures such as Alex Stamos, former Chief Security Officer at Facebook, Casey Ellis, founder of Bugcrowd, Jon Callas, former security lead at Apple, Paul Vixie, Dino Dai Zovi, former head of applied security engineering at Block, Katie Moussouris, founder of Luta Security, and Rachel Tobac, CEO of SocialProof Security. These experts highlight that Anthropic's models are essential for identifying vulnerabilities and securing software.

When launched in April, Mythos was presented as a powerful tool for vulnerability detection, initially limited to 50 companies, later expanded to 150 organizations across 15 countries. Last week, Anthropic introduced Fable, a public version of Mythos, equipped with strict safeguards to prevent malicious use. However, these safeguards were so stringent that they essentially blocked any requests related to cybersecurity.

Controversies Surrounding Model Capabilities

Anthropic indicated that the White House's decision might be linked to a report on a jailbreaking method for Fable. Katie Moussouris reviewed a non-public paper from Amazon researchers, which she claims does not demonstrate a genuine bypass of the safeguards. She asserts that the researchers simply used Fable to fix known vulnerabilities in open-source code.

The open letter calls for transparent and fair regulations, developed through a democratic process based on scientific research. The experts insist that these regulations should only be enforced to the extent necessary to ensure public safety. They also point out that the jailbreaking method could be replicated on OpenAI's GPT-5.5, Claude Opus 4.8, and Anthropic's Sonnet, as well as on Chinese models like Kimi 2.7.

Katie Moussouris stated in a blog post that the behavior described in the Amazon article cannot be significantly corrected, and any attempt would only weaken the model for defense. She emphasized that defenders must be able to ask the AI to fix bugs in a file, explain why the fix is important, and write tests that confirm the fix works. According to her, this is not a bypass of the safeguards, but the most valuable thing an AI model can do for defensive security: execute the search, fix, and test loop that defenders perform every day.