|
On June 9, Anthropic released Claude Fable 5, the most capable model it had ever made available to the public, and gave it away free to its paying tiers through June 22. On Friday, June 12, at 5:21 in the afternoon, the United States government told Anthropic to switch it off. The company complied within hours, taking down both Fable 5 and Mythos 5, its restricted cyber sibling. Other Claude models stayed up. The flagship, three days old, went dark.
The order was an export-control action citing national security, and it nominally targeted access “by any foreign national, whether inside or outside the United States.” Anthropic cannot tell in real time which of its users are foreign nationals, so to comply it had to turn the model off for everyone. An export control written to keep a model away from foreigners ended up keeping it away from Americans too, which is the kind of outcome you get when the enforcement tool predates the thing it is enforcing against.
The stated trigger was a jailbreak. By Anthropic’s account, the government produced only “verbal evidence of a ‘potential narrow, non-universal jailbreak’”, which amounted to prompting the model to read a codebase and point out the security flaws in it. OpenAI’s GPT-5.5 does this too, on request, today. Anthropic said it disagreed that a narrow potential jailbreak should be cause for recalling a model used by hundreds of millions of people. It is a reasonable objection, and it is also not obviously the whole story, because the most conspicuous fact about this recall is not the jailbreak. It is whose model got recalled.
Anthropic has spent the past two months in an open fight with the Pentagon, which designated the company a supply-chain risk, ordered a phase-out, and ended up in court even as the NSA kept using Anthropic’s models for its own cyber work. So when a national-security order arrives to switch off Anthropic’s flagship over a flaw-finding capability that its competitors share, the safety rationale is one available reading and a continuation of an existing grudge is another. It could be principled. It could be pettiness. From the outside the two are genuinely hard to tell apart, and that ambiguity is itself the story: Anthropic does not appear to be on good terms with the single regulator most able to end its products, and on Friday that regulator ended one.
The irony is hard to miss. Anthropic is the lab that made calling for AI regulation its brand. It warned that models were improving fast enough to be worth pausing over and pushed governments for tighter controls, and it is now the lab that has been regulated more bluntly than any other, its best model pulled by federal order within seventy-two hours of launch. Whether the advocacy has anything to do with the recall is unknowable. The pairing stands regardless. This, it turns out, is what regulation can look like when it arrives. Not a white paper or a panel or a thoughtful essay about timelines. A phone call at 5:21 on a Friday.
How that Friday went is itself in dispute, which tells you where the relationship stands. A Trump administration official said that during a flurry of calls the company claimed Dario Amodei was at a wellness retreat and unreachable, and called the conduct “recklessness.” A source close to Anthropic said there was no retreat and that executives were on the phone with the White House within fifteen minutes. Ashlee Vance, the Bloomberg journalist, was inside Anthropic’s headquarters reporting a story that day and said publicly there was no retreat and Amodei was right there. The government and the company cannot agree on whether the chief executive picked up the phone.
Every other AI lab is reading this closely, because every one of them is exposed to the same thing. They all ship models that can read code and find its flaws. They all depend on the same government for export rules, for contracts, and for the benefit of the doubt. What Friday demonstrated is that a frontier model can be switched off in an afternoon, worldwide, on a rationale the company disputes and outsiders cannot check, and that whether it happens to you may turn less on what your model does than on how your last few months with the government have gone.
|
