At 5:21 p.m. Eastern Time on Friday, Anthropic received a letter from Commerce Secretary Howard Lutnick invoking national-security export controls and barring any foreign national from accessing its two most capable models, Fable 5 and Mythos 5. The company had no mechanism to comply selectively and disabled both models worldwide within hours, including for its own non-citizen staff. Older releases such as Claude Opus 4.8 remain available.

It’s the first time Washington has applied export controls to an AI model itself rather than to the semiconductors used to train one. The semiconductor regime, built up since 2022, treated compute as the chokepoint. Friday’s order treats the weights as the weapon.

The Wall Street Journal first reported the letter’s contents, addressed to Anthropic chief executive Dario Amodei. Pentagon CIO Kirsten Davies posted on X that the action was about “prioritizing national security and the security of our warfighters.” Neither the Commerce Department nor the Pentagon responded to requests for comment.

Officials, according to Anthropic’s public statement, pointed to a jailbreak of Fable 5 that surfaces cybersecurity capabilities of the underlying Mythos system. Anthropic says it reviewed a demonstration and judged the technique narrow rather than universal, and noted that comparable capabilities “are able to be discovered” in other publicly available models. The company said the government “did not provide specific details” of the threat and added: “We believe this is a misunderstanding.” Anthropic, which earlier in June filed confidentially for a public listing, said it’s complying while seeking to restore access.

The diplomatic reception abroad has been sharp. Speaking in Ireland on Sunday before the G7 summit at Évian-les-Bains, Canadian Prime Minister Mark Carney said he had spent 45 minutes the previous evening discussing AI with French President Emmanuel Macron. “The situation we’re in collectively right now with Mythos and Fable is something that can happen with overreliance on certain models,” Carney said. “It is never a good idea to have one option.” British AI and Online Safety minister Kanishka Narayan wrote on X that the episode underscored why sovereign AI capacity was now a strategic question.

On Monday a group of cybersecurity researchers published an open letter calling the order “dangerous” and arguing that Washington had “taken the best models away from defenders” while equivalent capability sits on OpenAI’s GPT-5.5, Anthropic’s own Claude Opus 4.8, and China’s Kimi 2.7.

The export-control framework was designed to keep frontier capability inside American borders. Its first invocation against a model has done the opposite of reassure allies.

Sources