Trump administration bars foreign access to Anthropic's Mythos and Fable models

Anthropic took its most advanced AI models offline last week after the Trump administration imposed export controls restricting their use by foreign nationals, including the company's own non-U.S. employees. The move followed reports that Amazon researchers had identified a jailbreak allowing users to bypass safety features in the Fable model, raising national security concerns.

The restrictions applied to both Fable and Mythos, two frontier systems developed by the company. Anthropic said the vulnerability was narrow and not unique to its products, but it complied by removing access entirely rather than attempting to limit distribution. Administration officials described the decision as a response to warnings from multiple companies and a failure by Anthropic to address risks that its own leadership had previously highlighted in public statements about AI dangers.

Sources familiar with the discussions said the episode reflected ongoing friction between Anthropic and the administration. Officials argued that the company had received approval to deploy Fable yet later resisted stronger safeguards, creating a perception that its safety commitments applied unevenly. One official noted that Anthropic appeared to have chosen the riskier path at several decision points, including how it handled early information about potential misuse.

The company's public emphasis on AI risks has been a central part of its positioning since its founding. Chief executive Dario Amodei has repeatedly warned of catastrophic outcomes if advanced systems are not properly controlled. Administration officials viewed the gap between those warnings and the company's response to the specific jailbreak as inconsistent, particularly as other firms raised alarms before the export controls were issued.

The restrictions come at a moment when U.S. policy on frontier AI is still taking shape. Export controls represent one of the few existing tools for limiting access to the most capable models, but they require coordination between companies and regulators that has proven difficult. Anthropic maintains it followed government guidance on deployment, while officials contend the company did not adequately anticipate or mitigate downstream vulnerabilities.

The episode also carries competitive implications. Chinese AI developers, including Zhipu, moved quickly to release new open-source models without usage restrictions, and their shares rose sharply after the U.S. action. Investors appear to be pricing in the possibility that tighter American controls could shift demand toward alternatives less constrained by export rules.

For Anthropic, the immediate result is the loss of a major product line while it works through the regulatory issues. The episode underscores the challenge of aligning private-sector development timelines with government expectations around national security, especially when the technology in question is both commercially valuable and potentially dual-use. How the company and the administration resolve the underlying communication gaps will likely influence future attempts to govern advanced AI systems.