Export-control order against Anthropic exposes AI policy gap
Anthropic blocked its Fable 5 and Mythos 5 models after a U.S. directive that experts say tests unclear rules for hosted AI access.
By Hana Yoshida · Markets Reporter
3 min read
Anthropic has been trying to restore access to its newest AI models after the Trump administration ordered the company to cut off foreign nationals, The Verge reported. The episode matters because experts told The Verge that applying export controls to access for hosted AI models appears to push into unsettled legal territory.
The order affected Fable 5 and Mythos 5, according to The Verge, and covered foreign nationals both outside the United States and inside the country. The Verge reported that Anthropic blocked the models for all users, including its own employees, as it worked to comply.
Anthropic said on its website that the government pointed to “national security authorities” and issued what the company described as an export-control directive. The company also said the government had raised concerns about a possible jailbreak linked to groups connected to China, while Anthropic said that issue did not let users evade all of its safeguards.
The Trump administration has not publicly detailed the legal basis for the order, according to The Verge. Anthropic did not respond to The Verge’s request for comment.
Why the order is unusual
Export controls have long been used for goods that can cross borders, including weapons, equipment and other hardware. The rules later expanded to cover less tangible items such as software, source code and technical data, The Verge reported.
AI models have already drawn attention from U.S. policymakers under that framework. The Biden administration moved to regulate AI model weights, the data that allows a model to run elsewhere, but the Trump administration later rescinded that approach, according to The Verge.
The Anthropic case differs because Fable 5 and Mythos 5 remain on Anthropic’s servers, The Verge reported. Users receive outputs from a chatbot rather than a copy of the model, its code or its weights.
That leaves several possible theories for what the government viewed as an export. The Verge reported that it could be the model’s responses, access to the service itself, or some other category not yet clearly defined in current rules. Remote access to cloud services is already a gap lawmakers are trying to address through legislation moving through the Senate, according to The Verge.
Experts warn of uncertainty
Hanna Dohmen, a senior research analyst at Georgetown University’s Center for Security and Emerging Technology, told The Verge that without the exact wording of the order, it remains unclear whether the directive stretches existing rules. She said it appears to be the first use of U.S. export controls to restrict access to an AI model in this manner.
Andrew Reddie, a professor at UC Berkeley’s Goldman School of Public Policy, told The Verge that export-control and arms-control systems give the government broad room to restrict certain goods. He said shifting expectations from one administration to another have made it harder for AI companies to know what duties they must meet.
The directive raises questions for other frontier AI companies if Anthropic was targeted because the models were unusually capable, The Verge reported. If the issue was safeguards, experts told The Verge that companies need clearer standards for what counts as adequate protection.
Reddie also warned The Verge that requiring models to be impossible to jailbreak would leave the United States without AI models. Experts told The Verge that unclear, case-by-case intervention could also strengthen arguments by foreign governments and companies that they should not rely on U.S. firms for strategically important AI systems.
This story draws on original reporting from The Verge.