At 5:21 on Friday afternoon, Anthropic received a letter, and by that evening two of its models had vanished for everyone on the planet. Claude Fable 5 and Mythos 5, gone. The instrument was an export-control directive, the kind of authority built to keep advanced chips and weapons designs out of hostile hands. This time the administration pointed it at a chatbot.

The directive bars foreign nationals from using either model, including Anthropic's own foreign staff inside the United States. On paper that sounds narrow. In practice a company cannot sort its users by passport in real time, so the only way to comply is to switch the models off for everybody. A rule written to stop technology crossing a border became a global kill switch, and it worked in a single afternoon.

The stated reason is thin. The government believes someone found a way to jailbreak Fable 5, and that is most of what we have been told, because the letter carried no technical specifics. Anthropic, which actually saw the demonstration, says the technique exposed a small, already-known software vulnerability, the precise category of flaw that Fable's safety design was built to catch, and nothing a person couldn't already coax out of GPT-5.5. That last detail is the tell. The exact capability the administration judged too dangerous for Fable, which Anthropic says it had deployed to hundreds of millions of people, sits right now, unrecalled, inside a competitor's product. One model dies over a weekend; its functional twin stays online. That is not how a government acts when it has found a weapon. It is how one acts when it has found a lever.

This is the first time Washington has forced a commercial AI product offline, and the manner of it should worry people who have never touched Anthropic. The whole thing took three days from launch to recall, ran on what the company describes as verbal evidence, and arrived with no published finding and no chance to contest the switch before it was thrown. Calling it disproportionate is too polite. A government that can erase a widely used service over a weekend, citing a secret it will not show anyone, has found a tool far more useful to it than any jailbreak.

The lesson most people are drawing is about resilience, about not leaning on a single vendor. That is sensible, and it misses the point. The dependency that failed on Friday was not on Anthropic. It was on an administration choosing not to use a power it turns out to hold. Export-control law hands the executive enormous discretion and almost no duty to explain itself, and it has now been aimed at software that ordinary people had open in a browser tab. The company that spent the spring asking for a verifiable brake on frontier AI just found out what a brake feels like when someone else holds it. The rest of us get to wonder which model the off switch finds next.

Sources: