Anthropic called Mythos a cyberweapon, government agreed

At 5:21 p.m. Eastern on June 12, 2026, Anthropic received a letter from the Commerce Department. The company says it was given ninety minutes to pull its newest model. By nightfall, it had taken both of its most powerful AI systems offline for every customer worldwide.

Claude Fable 5 and Claude Mythos 5 had launched three days earlier. Mythos went to a small set of vetted security partners. Fable, the same model under a layer of safeguards, went to the public. It was the first time a leading AI company had pulled a publicly deployed model because the government told it to.

The mechanism was export control. Anthropic's statement said the government, citing national security authorities, barred any foreign national from using the models, whether outside or inside the United States. The order reached Anthropic's own foreign-born engineers. It reached foreign nationals working at American desks.

The letter came from Commerce Secretary Howard Lutnick and went to CEO Dario Amodei. The legal theory appears to be the deemed export rule, which treats showing controlled technology to a foreign person as an export to their home country. It was written for semiconductor and encryption applications. Chris McGuire of the Council on Foreign Relations, who reviewed the move, allowed that targeted controls could be prudent but called applying deemed export to foreign nationals inside the country "just absurd," and said it would force the model out of distribution, which is what happened.

Anthropic could not filter by citizenship in real time across its accounts and APIs, so it shut down the models for everyone. New requests now route to Claude Opus 4.8. Every other Claude model stayed up.

Two Accounts

Anthropic complied and disagreed in the same statement. The letter gave no detail; it said. It was understood that someone had found a way to jailbreak Fable 5: ask the model to read a codebase and flag software flaws. The vulnerabilities that surfaced were minor and already known, and other public models find the same ones with no jailbreak at all, including OpenAI's GPT-5.5.

No tester had found a universal jailbreak, one that broadly defeats the safeguards. Perfect resistance is not possible for any model, Anthropic argued, and a narrow flaw does not justify recalling a product used by hundreds of millions. Apply that standard to everyone, the company said, and you stop frontier AI deployment across the industry.

The White House told it differently, using Anthropic's own words.

David Sacks, who co-chairs the President's Council of Advisors on Science and Technology, gave the administration's version on X. Fable is Mythos with guardrails. When the guardrails fail, Mythos and its cyber capabilities reach people who should not have them. Anthropic had spent months calling Mythos a cyberweapon and asking Washington to regulate it as one. A trusted partner testing Fable found a jailbreak. The administration asked Amodei to fix it or pull the model.

In Sacks's words, "Dario refused."

Anthropic prioritized its consumer product over safety, Sacks wrote, calling that at odds with everything the company says it stands for. The export control followed reluctantly. Patch the flaw, and it lifts.

One day before the shutdown, Amodei had published an essay, "Policy on the AI Exponential," arguing that the government should hold legal authority to block or reverse the release of frontier models that fail safety testing. He compared it to the FAA grounding of unsafe aircraft. When the government blocked his, Anthropic called the process improper.

Two facts sit underneath both accounts.

The partner who reported the jailbreak was, by multiple accounts, Amazon: Anthropic's largest investor and its main cloud host. Amazon CEO Andy Jassy raised it with senior officials after his researchers prompted the model into producing restricted attack information. Whether Amazon tested Fable on its own or at the government's request is not known.

The other is China. Semafor reported the White House acted partly on suspicion that a China-linked group had reached Mythos. Anthropic told the outlet that the White House never raised Chinese access in its talks and that it blocks access to its products from inside China. The NSA had been given Mythos 5 for offensive cyber operations, which leaves the order pointing partly at the government's own supply.

The capability is not in dispute. In Mythos Preview's first month, Mozilla found and fixed 271 vulnerabilities in Firefox using the model, more than ten times what Claude had found earlier. What people argue about is the guardrails. Peter Girnus, a software threat researcher, put it plainly on X: "If you describe your product as a munition in every press release, eventually a government takes you at your word."

Before this

The shutdown landed in the middle of a fight already in court.

In March, the Pentagon labeled Anthropic a supply chain risk, a designation built for foreign adversaries, after the company refused to let the military use Claude for mass surveillance or fully autonomous weapons. Anthropic sued. A California judge blocked the label. The D.C. Circuit let it stand and heard arguments in May, with the panel split. Sacks says the Fable order has nothing to do with any of it. Anthropic has not said the same.