U.S. lifts controls on Anthropic models but policy gaps remain

The U.S. government has rolled back export controls on Anthropic’s Mythos and Fable models after restrictions forced the company to disable both systems for all users. The reversal restores access to Fable, which can carry out lengthy tasks autonomously, and Mythos, which cyber defenders had hoped to use to identify and patch security flaws before attackers gain comparable capabilities.

The decision leaves U.S. AI policy in an uncertain state. Washington is effectively operating a licensing regime for frontier AI models while denying that such a regime exists, with rules that appear opaque and improvised. Reports suggest the government is working with leading AI labs on voluntary cybersecurity standards that could give companies clearer expectations before releasing advanced models.

Anthropic is also developing a shared framework with the U.S. government, Amazon, Microsoft, Google, and other Glasswing partners to assess the risks posed by jailbreaks. The export control episode has already pushed some potential customers, especially in Europe, to question whether they can rely on American frontier models for essential uses and to consider fallback options, including open source alternatives.

More capable open source models create a separate policy dilemma because guardrails and classifiers can be removed or bypassed when model weights are available. Intelligence agencies have warned of an imminent cyber threat from advanced AI models, while OpenAI’s Sam Altman has renewed calls for a U.S.-led international governance framework to coordinate standards and access among Western governments.