After the U.S. government requested Anthropic to restrict the export of two advanced models, Fable and Mythos, a group of cybersecurity professionals publicly opposed the decision, arguing that the restrictions undermine the ability of defenders to use advanced AI to discover vulnerabilities, fix flaws, and strengthen software.
The open letter has been signed by 76 cybersecurity experts, including former Facebook security chief Alex Stamos, Bugcrowd founder Casey Ellis, cryptographer Jon Callas, computer scientist Paul Vixie, and Luta Security founder Katie Moussouris. The letter states that, in the face of rapidly advancing adversary capabilities, removing the strongest capabilities from defenders is a dangerous approach.
Global access suspended following government order
Anthropic stated that the U.S. government issued an export restriction order last week citing national security, without specifying further grounds. The company subsequently suspended global users' access to Fable and Mythos.
Mythos was launched in preview form in April this year. At the time, Anthropic stated that the model was highly capable at identifying security vulnerabilities, necessitating strict access controls to prevent misuse by malicious hackers or foreign adversaries. Initially, only about 50 companies were granted access, and this was later expanded to approximately 150 institutions across 15 countries.
The dispute centers on the so-called bypass methods.
Anthropic also noted that the White House's restriction order may be related to a report claiming that a method exists to bypass Fable's limitations, thereby unlocking capabilities approaching the Mythos level.
Katie Moussouris stated that this unpublished research, conducted by Amazon researchers, was reviewed by her personally. She believes the paper does not demonstrate a true "jailbreak" method. According to her, the researchers simply asked the model to repair open-source code containing publicly known vulnerabilities and artificially inserted vulnerabilities, rather than successfully bypassing the model’s existing safeguards.
She noted that such operations are part of the security team's standard workflow, involving identifying issues, fixing vulnerabilities, and verifying that patches are effective. Viewing this capability as a way to bypass restrictions would only diminish the model's value in defense scenarios.
The signatories stated that the issue is not unique to Anthropic.
The open letter also stated that these methods are not limited to Anthropic models; similar issues may also occur with OpenAI’s GPT-5.5, Anthropic’s publicly available Claude Opus 4.8 and Sonnet, and even China’s Kimi 2.7 model.
The signatories also called on the U.S. government to increase transparency in developing relevant regulations, advance the process through more open procedures, and minimize restrictions based on research from the industry and academia.