Washington has switched off two of the most powerful AI systems ever built — for the entire world, on a few hours’ notice. The official reason is national security. The fuller story is stranger, and more revealing.
At roughly the same moment one arm of the United States government decided that an artificial-intelligence system called Mythos was too dangerous to let anyone abroad touch, another arm of that same government was reportedly using it to break into the computer networks of America’s enemies. A technology treated as a hazard and a weapon at once: that contradiction sits at the centre of the most extraordinary clash yet between Washington and the company widely seen as the most safety-obsessed in the AI business.
The trigger came on Friday evening. Anthropic, the San Francisco firm behind the Claude chatbot, received a letter from the US Commerce Department, and by the end of the night it had pulled the plug on its two newest and most capable systems, Fable 5 and Mythos 5, not only in the United States but everywhere on earth. The order did not, strictly speaking, tell the company to shut anything down. It said only that Anthropic could no longer give access to any “foreign national” — anyone who is not an American citizen, whether they sit in Bengaluru or in a cubicle in Manhattan. Faced with a rule that sweeping, the company decided the only way to obey it was to switch the systems off for everyone. Its older models, including the one called Opus 4.8, kept running.
What makes the move so striking is the instrument the government reached for. There is no dedicated American law for policing AI, so Commerce used export controls instead — the decades-old machinery built to keep missile parts and advanced weaponry out of the wrong hands, the same regime Washington has used to deny cutting-edge computer chips to China. Under that law, even showing sensitive technology to a foreigner standing on US soil counts as an export. An order ostensibly about foreign access thus became, in practice, a global off-switch, and a commercial product went dark over a weekend.
The stated fear is a “jailbreak”, the term for a trick that coaxes an AI past its own safety rules. According to Axios, which first reported the letter, an administration official said the action followed a claim by a rival company that it had found a way to defeat Mythos’s guardrails, and that the systems would stay frozen until the government had hardened its own cyber defences, perhaps over several weeks. Anthropic tells a far more deflating version of the same events. It says it studied the technique and found it surfaced only minor, already-known weaknesses, the kind ordinary AI systems, including OpenAI’s GPT-5.5, can turn up without any trickery at all. No one, it maintains, has discovered a master key able to unlock the model’s defences wholesale. The company calls the whole affair a misunderstanding, and warns that if a single narrow flaw can justify recalling a mass-market product, no AI firm could ever release anything with confidence.
It is worth being honest about how little an outsider can actually verify. By Anthropic’s own account the letter did not name the specific danger, and the government has shown its evidence to no one. Anthropic has every commercial incentive to play the threat down; the administration has its own reasons to play it up. Both accounts cannot be fully true, and for the moment neither can be tested.
Which is why the history matters more than the headline. The very capability the United States has just judged too dangerous to export is the capability it spent earlier this year trying to punish Anthropic for guarding too jealously. In July 2025 the company signed a two-year Pentagon contract worth up to $200 million and became the first AI lab to run its models on classified military networks. The relationship curdled when Anthropic refused to drop two conditions it had written into the deal: that its AI would not be used for mass surveillance of Americans, nor to operate fully autonomous weapons. The Pentagon wanted access for “all lawful purposes” and would not take no for an answer.
When the talks broke down in late February, President Trump ordered federal agencies to stop using Anthropic’s tools and Defence Secretary Pete Hegseth branded the firm a “supply chain risk” — a label devised for companies such as China’s Huawei and never before pinned on an American business. Anthropic sued. In late March a federal judge granted an injunction that partly blocked the blacklist, finding the company likely to prove the whole exercise was unlawful retaliation for its outspoken stance on AI safety; part of the designation survived, and the case grinds on. US President Donald Trump, for his part, called the company “Leftwing nut jobs”.
Then comes the detail that gives the story its real shape. Even as the Pentagon was casting Anthropic out, the Financial Times reported that the National Security Agency had quietly embedded around half a dozen of the company’s engineers to help it run Mythos, reportedly with offensive operations against adversaries such as China and Iran in mind. The paper was careful to add that it was unclear whether the model was being used in any live attack. Still, the ledger reads plainly enough: too risky for the government to buy, too risky to sell to the world, and yet useful enough that America’s most secretive spy agency wanted its makers in the room.
Seen together, those three positions are less contradictory than they look. They are the conduct of a state that has decided artificial intelligence is no longer ordinary software but a strategic weapon — something to be hoarded, denied to rivals and turned on enemies, much like enriched uranium or a fighter jet. That reframing, not the weekend’s outage, is the real news, and it will outlast this particular quarrel.
For everyone outside the United States the immediate consequence is blunt. Every developer, student and business beyond America’s borders, including the vast number of them in India, has lost two of the finest AI systems yet made, with no say in the decision and no firm idea when, or whether, access returns. For investors the episode plants a colder thought. An American AI company’s flagship product can now disappear on a government’s say-so, overnight, a political risk that no balance sheet quite knows how to price.
Whether Friday’s order was a sober act of security or simply the latest blow in a feud that has already reached the courts, no one outside a few classified rooms can yet say. What is no longer in question is the larger truth the affair exposes. The most powerful AI being built in America has become valuable enough that its own government wants it in hand, and frightening enough that it would prefer no one else had it at all. Both things are true at once, and that is the uneasy place the technology, and the rest of us, have now arrived at.