Friday evening, a letter shuts down an AI
On 12 June 2026, at 17:21 (New York time), Anthropic receives a letter from the American government. A few hours later, two of its most powerful artificial intelligences, Fable 5 and Mythos 5, display the same message to hundreds of millions of people: "temporarily unavailable".
This is not an outage. This is not a bug. This is an order from the state.
And to our knowledge, it is the first time in history that a government has shut down a consumer AI that was already in service. That is why everyone is talking about it.
The weapon used: a rule designed for China, aimed at an American company
The decision comes from the US Department of Commerce. The tool employed is export control. Normally, it is used to stop sensitive technologies (chips, military secrets) from falling into the hands of a rival country.
There is a rather mad American rule in all this. In the United States, showing a sensitive technology to a foreigner, even if they are on American soil, counts as an "export" to their country. Imagine a secret plan: letting a foreign visitor read it in your living room is, in the eyes of the law, the same as having sent it abroad. That is exactly the logic that was applied to Fable 5.
Except that here, the target is not an adversary. It is an American company selling a product to the general public. And the ban even covers Anthropic's foreign employees. As a result, to be sure of complying with the rule, the company unplugged everything. Its other models, including the recent Claude Opus 4.8, are still working normally.
The accusation: a "jailbreak"... that other AIs can also do
What exactly is the state accusing these models of? The letter does not say clearly. According to Anthropic, the government believes it has found a way to get around Fable 5's safeguards. In jargon, this is called a jailbreak: a trick to make the AI say or do something it is supposed to refuse.
But Anthropic's version is far less dramatic. The company says it has looked at the file, and that the flaw essentially consists of asking the model to read a piece of computer code and spot security vulnerabilities in it. That is to say... the daily work of any cybersecurity team.
The detail that hurts for the government: according to Anthropic, other consumer AIs, including OpenAI's GPT-5.5, can do exactly the same thing. And nobody has unplugged them. So if the flaw exists, there is nothing unique or secret about it.
A universal jailbreak = the master key that opens every door of the model at once. The doomsday scenario.
A limited jailbreak = a trick that unlocks one small thing, in one specific case.
Anthropic maintains that nobody has ever found a master key for Fable 5, and that at worst it is being criticised for a small, limited trick.
The most ironic part: Anthropic actually sells security
This whole story has a taste of irony. Fable 5 was launched with protections so strict that some users found them downright too severe. Before release, the company says it had the model tested for thousands of hours, with the US government itself and the British AI safety body.
Its philosophy: start from the premise that no AI is 100% tamper-proof, and therefore stack up the barriers. Make workarounds either tiny or prohibitively expensive to carry out, and monitor continuously so it can react fast. The company swears it has never seen a single case of a workaround causing real harm.
So its position is clear: shutting down a product used by hundreds of millions of people over a minor flaw is, it argues, opening a Pandora's box. If that rule were applied across the industry, no new AI could ever be released again. Anthropic accepts that the state should be able to block a genuinely dangerous product, but it wants a clear, fair procedure based on technical evidence. Which, it says, was not the case here.
The real heart of the matter: a fight that has been going on since March
To understand, you have to zoom out. This is not a bolt from the blue. It is the latest episode in an open conflict between Anthropic and the Trump administration.
Back to March 2026. The Department of Defense classifies Anthropic as a "national security risk", a label usually reserved for enemy countries. The reason? Anthropic refused to let its AI be used for two things: fully autonomous weapons and mass surveillance of citizens.
Anthropic took the government to court. And several judges were harsh on the state: one reckoned the company appeared to have been "punished", another spoke of a "spectacular overreach". The case is still ongoing today.
Why it concerns us, even from France
"An American affair, who cares"? Not so fast. Here is what it puts on the table, for us too.
It is a global precedent. The first time a state has switched off a mainstream AI in operation. How it is resolved will serve as a model for every government, Europe included, the day they want to intervene on an AI.
It exposes our dependence. Overnight, people and businesses around the world lost their working tool, without warning. It is a reminder that our access to these AIs hangs by a thread. From now on, by a minister's signature.
It raises the real question. Who decides that an AI is "too dangerous"? On what criteria? With what evidence? When even the company most obsessed with safety gets cut off in the name of safety, the debate becomes far more complicated than a simple "for or against regulation".
And it calls the logic into question. If the same capability exists at competitors that are left untouched, unplugging a single player does not remove the risk. It moves it, that is all.
And now?
Anthropic promises to say more very soon and says it is doing everything to restore access, speaking of a "misunderstanding". The government, for its part, has so far given only verbal explanations about a limited flaw. Between a company crying error and a state invoking national security without revealing everything, the truth is surely hidden in the parts of the file we cannot see.
One thing is certain, though. The day a minister can switch off an artificial intelligence with a stroke of the pen has just arrived. The question is no longer whether the state will intrude into AI. It is according to what rules, and who will write them.