At 5:21pm ET on June 12, 2026, Anthropic received a directive from the US Commerce Department ordering it to suspend access to Fable 5 and Mythos 5 for foreign nationals. Anthropic suspended both models for everyone, not just foreign users, because targeted geographic compliance wasn’t feasible to implement cleanly. As of this writing, both models are offline.
Audio
Listen to this article
A 2-minute audio overview of this article, narrated by our robot.
This is the first known instance of the US government using export control authority to force frontier AI models offline mid-deployment.
What Happened
Commerce Secretary Howard Lutnick sent the letter directly to Anthropic CEO Dario Amodei. The trigger was a report from an unnamed third-party company claiming it had jailbroken Mythos 5. The administration moved to restrict foreign access on national security grounds.
The jailbreak in question: prompting the model to read a codebase and fix software bugs. The technique exposed what Anthropic describes as “a small number of previously known, minor vulnerabilities.”
Anthropic’s response in their public statement is pointed. The capability being cited as a national security threat is one that security professionals use daily with GPT-5.5 and other frontier models for defensive work. Restricting Fable 5 and Mythos 5 doesn’t remove the capability from the world. It just removes two models from the market while every comparable alternative remains available.
Anthropic says they reviewed the demonstrated jailbreak, found it narrow and non-universal, dispute the standard being applied, and are complying while disagreeing. They’re working to restore access.
The Context That Matters
This didn’t come from nowhere. The Trump administration had previously attempted to block Fable 5’s release entirely, but that effort was unsuccessful. The export control directive is a second attempt to constrain the models, this time with a specific technical justification.
Whether that justification holds is a different question. The Commerce Department hasn’t published the letter or the evidence. The identity of the company that reported the jailbreak hasn’t been disclosed. The specific statutory authority invoked hasn’t been confirmed in public reporting. What we know is that a third party made a claim, the administration acted on it within hours, and Anthropic is now running without its two most capable models while it pushes back through whatever channel it has available.
I Tried to Use It
While writing this, I tried to switch the session drafting it over to Fable 5. The request failed: “There’s an issue with the selected model (claude-fable-5). It may not exist or you may not have access to it.” The session fell back to another model and kept going.
That is the takedown, observed firsthand rather than read off a press release. A few hours earlier, Fable 5 was the most capable model Anthropic shipped. Now a routine API call for it returns the same error you’d get for a model that was never released. The directive isn’t a policy memo working its way through a queue. It’s live, at the API layer, right now.
There’s an irony worth sitting with. The capability the government cited as the threat is reading code and fixing bugs, exactly the work this model was being asked to do when the request bounced. The thing that got Fable 5 pulled is the thing it does best, and the thing every other frontier model on the market still does freely.
Why This Is Bigger Than One Incident
The precedent question is what matters most here. If the Commerce Department can compel an AI lab to take models offline globally based on an undisclosed third-party jailbreak report, with no public evidence, no stated standard, and no appeal window before the suspension takes effect, then every frontier lab is operating in a different environment than it was yesterday.
The jailbreak cited is a code-reading capability. That bar, applied consistently, would catch capabilities that exist in every major frontier model currently deployed.
Anthropic is working to restore access. Whether the administration lets them is an open question. The broader question, what standard if any governs when the government can pull a model, has no answer yet.