At 5:21pm Eastern on Friday, an export control directive from Commerce Secretary Howard Lutnick landed in Anthropic’s inbox. By the end of the evening, Claude Fable 5 and Mythos 5 were dark for every user on Earth. NBC News calls it the first time a leading AI company has pulled a deployed model offline at federal instruction, and the framing matters: this isn’t a recall, it’s a precedent.

The letter, drafted with help from the Bureau of Industry and Security, barred foreign nationals from accessing either model. Anthropic, unable to filter users in real time, opted for the global killswitch. Every other Claude model, including Opus 4.8, stayed live.

Fable 5 had launched Tuesday. Three days.

Mythos 5 is the more interesting object here. Previewed in April and distributed only through Project Glasswing to roughly 50 vetted partners (Amazon, Apple, Google, Microsoft, CrowdStrike) for defensive security work, it was the underlying model; Fable was Mythos with classifier-based guardrails clamped over cybersecurity and biology. Vals AI clocked Fable 5 as the most capable model publicly available. That ranking lasted seventy-two hours.

The directive didn’t spell out the national security concern. Anthropic’s read, stated publicly, is that the government believes someone jailbroke Fable 5 and shared a demonstration, one the company says surfaced only previously-known minor vulnerabilities reproducible on other public models including OpenAI’s GPT-5.5. The company’s pushback was unusually direct: “We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.”

Cybersecurity researcher Peter Girnus posted the obvious read on X, Anthropic’s own safety rhetoric, the years of warning that frontier models approach weapons-grade capability, invited the regulator to take that framing literally and classify the models as munitions.

The timing compounds. Anthropic confidentially filed for an IPO earlier this month, and is already litigating a separate Defense Department supply-chain blacklisting. A company built on the proposition that it understands AI risk better than its rivals is now discovering what happens when the state agrees, then acts on it. The safety-pilled lab has spent years arguing the models are dangerous. Lutnick read the memo.

Sources