The US government just banned Claude

Anthropic spent the first half of this year telling everyone their best models were too dangerous to ship. On Friday evening, the US government took them at their word, and then went one step further. It ordered Claude switched off.

At 5:21pm Eastern on the 12th, Anthropic received a directive citing national security authorities. They had to disable Fable 5 and Mythos 5, their two most capable models, for every customer “to ensure compliance.” Not throttle. Not gate. Off. The rest of the Claude lineup stays up, so this is not the end of the company. But the flagship, the one people built real workflows on, went dark on a few hours’ notice.

I have a small weekend ritual. When a new model lands, I hand it a passion project, something with no client attached and no deadline, just to find where the ceiling is. Opus 4.8 got my World Cup predictor. Fable 5 was next in line. I had the weekend set aside for it. I never got to open the tab. They killed it before I could type a single prompt.

I wrote back in April that “too dangerous to release” read more like a press release than a confession. This is the part where it stops being a marketing line and becomes someone else’s decision.

A government can now reach into a product you pay for and turn it off by Friday evening. That is not a safety feature. That is a supply chain you do not control.

The reason is the part that should bother you. The government believed it had found a way to jailbreak one of the models. Anthropic looked at the technique and described what it actually was: asking the model to “read a specific codebase and fix any software flaws.” That is not an exploit. That is the literal job. It is the thing every developer using these models does on a Tuesday. Anthropic said as much, that “we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model,” and noted that other models, OpenAI’s GPT-5.5 included, can do the same thing. The most capable coding model on the market got pulled because it was good at coding.

What you actually rented is the second part. If your team standardised on Fable 5, your Monday standup now has a hole in it where a tool used to be. You did not get a deprecation notice. You did not get a migration window. You got a kill switch flipped by people who do not use your sprint board. The pragmatic lesson here is old and boring and true. When the thing you depend on lives on someone else’s servers under someone else’s licence, you are renting access, and access can be revoked by people you will never meet. Keep a fallback model you can switch to in an afternoon, because the afternoon may not be yours to choose.

The funny part, if you squint, is that Anthropic and the government finally agree on something. One side said the model was too powerful to release. The other side said it was too powerful to keep running. Somewhere in the middle is a developer who just wanted to fix a bug.

If you want Anthropic’s own account of the directive, their statement is here. It is short, and unusually blunt for a company statement.