U.S. Commerce Department Halts Distribution of Anthropic's Fable 5 Model After Security Concerns

The U.S. Commerce Department announced on Friday that it is applying national-security export controls to prevent Anthropic from distributing its Fable 5 model and the underlying Mythos 5 system to foreign nationals, including non-citizen employees in the United States. The action follows a report that Amazon researchers discovered a prompt that could bypass the model’s safeguards and reveal information about cyber-attacks.

Amazon’s chief executive raised the issue with senior administration officials after the test, and the company said it often advises governments on potential security risks but does not disclose discussion details. Unnamed sources cited by media outlets indicated that the government suspected a Chinese-linked group had already exploited the vulnerability, though the evidence for that claim was not made public.

Anthropic’s chief executive, Dario Amodei, engaged in multiple calls with White House officials, arguing that the identified bypass was limited rather than a full jailbreak. According to a source familiar with the company, Anthropic received a 90-minute notice to withdraw the model and was not given prior warning of a national-security threat. The company subsequently disabled both Fable 5 and Mythos 5 for all users.

The export-control measure is the first time the United States has halted a widely available commercial AI model. An administration official told a news outlet that other AI models on the market are not viewed as posing the same level of risk because they have not reached the capability level of Mythos.

The decision escalates an ongoing dispute between Anthropic and the government that began earlier this year when the company was labeled a “supply chain risk” for Pentagon contractors after refusing contract terms that would allow unrestricted use of its models. Anthropic is contesting that designation in court.

White House AI adviser David Sacks said a trusted partner identified the jailbreak and that the administration asked Anthropic to fix the issue or withdraw the model. He added that the export control was issued reluctantly and that the government hopes the model can be reinstated after remediation.

The move has prompted reactions from European officials, who cite it as evidence for the need of “sovereign AI” – national control over AI models, infrastructure, and data. Former French prime minister Édouard Philippe and other European politicians warned that reliance on foreign-controlled AI could leave critical services vulnerable to sudden shutdowns.