Anthropic disabled its two most capable models, Fable 5 and Mythos 5, on June 12 after the US government handed it an export control directive citing national security. The order, which arrived at 5:21pm ET, bars any foreign national from accessing the models. Anthropic says it had no clean way to comply except to shut both off for everyone, US customers included.
The company laid out its version of events in a public statement Friday evening. Every other Anthropic model stays online.
What the government actually objected to
The letter didn't spell out the concern. Anthropic's read is that someone showed the government a way to jailbreak Fable 5, and the government found it alarming enough to act. Commerce Secretary Howard Lutnick sent the letter directly to CEO Dario Amodei, according to Axios reporting, after a rival company claimed it had cracked Mythos.
Here's where it gets thin. Anthropic says it reviewed a demonstration of the technique and what it found was a handful of previously known, minor vulnerabilities. The kind, the company argues, that other public models surface on their own without any bypass at all. The jailbreak in question apparently amounts to asking the model to read a codebase and fix software flaws, which is, you know, what these tools do all day for the people defending those systems.
The precedent problem
Anthropic is complying. It's also clearly annoyed. The company says a narrow, non-universal jailbreak shouldn't be grounds for recalling a model already deployed to hundreds of millions of people, and that applying this bar across the board would freeze frontier deployments industry-wide.
"We believe this is a misunderstanding and are working to restore access as soon as possible." Which is the diplomatic framing. The sharper line in the same statement is that the action doesn't meet the standards Anthropic has publicly argued any government deployment block should meet: transparent, fair, grounded in technical facts.
The company points to its own safeguards as evidence the move is overkill. It says it red-teamed Fable for thousands of hours before launch alongside the US government, the UK AISI, and outside groups, and that no one has found a universal jailbreak yet. Anthropic also notes it required 30-day data retention on Fable specifically to catch and shut down attacks, a policy it concedes cost it customers.
Worth a grain of salt: this is all Anthropic's account of a finding it hasn't seen in full. The company says it validated the capability against OpenAI's GPT-5.5 and found the same thing available there. We haven't seen the government's underlying report either.
Why this one matters
This looks like the first time a leading AI company has taken a publicly deployed model offline because the federal government told it to. Not a voluntary pause, not a pre-launch negotiation. A directive after the fact, on a commercial product in active use.
The administration had reportedly tried to get Anthropic to delay the launch and didn't succeed, which is what set up the export-control route. Both models trace back to Mythos Preview, the security-research system Anthropic ran through its Project Glasswing partners before this month's launch.
Anthropic says it will share more over the next 24 hours. Watch for formal documentation from Commerce and any narrowing of the directive.
