For the first time, the United States government has switched off frontier artificial intelligence and forced the world to go without it. Two of the most capable AI systems ever built, Claude Fable 5 and Claude Mythos 5, went dark last week. Not just in China or Iran. A researcher in London, a developer in Tokyo, an entire company in Berlin, all cut off at once, all treated as equally dangerous.
A letter reached Anthropic at 5.21 on a Friday afternoon from Commerce Secretary Howard Lutnick, citing national security authorities. It told the company to suspend access for every foreign national, anywhere in the world. There is no switch that lets in Americans and keeps out everyone else. To comply, Anthropic said it had to take the models down for all users at once.
On a commercial chatbot, America’s allies were handed the same blackout as China, an authoritarian rival
The precedent should alarm anyone who believes a serious government acts through evidence and law, not panic. Strip away the national security language and we are left with a new and dangerous precedent – the government’s conviction that it can switch off all software, anywhere, by writing a letter.
I have spent a career on the House Foreign Affairs Committee and as a correspondent who covered China, arguing that the West is too casual about handing its hardest technology to its adversaries. I take the cyber risk of these models seriously.
So take it from this hawk: this was not how you act on a real threat. The government moved in hours, on evidence that Anthropic disputes, and it reached for a Cold War export doctrine and stretched it to breaking point.
Anthropic is an awkward company to defend. For years now, it has told everyone AI models were dangerous. It built its brand on the warning, pressed Washington to take the threat seriously and got its wish. Dario Amodei, its chief executive, has said the world underestimates how good and how bad this technology could become. Now the government has turned that fear back on Anthropic, and not for the first time. The Pentagon had already branded the company a national-security risk after Anthropic refused to let its AI be used for autonomous weapons and mass surveillance, a fight a federal judge said looked like an “attempt to cripple” an American business.
Until this spring, Anthropic kept its most powerful model, Mythos, away from the public, releasing the model since April only to vetted corporate and research partners on the grounds that its cyber capability was too dangerous for open sale. Then, on 9 June, it released a public version, Claude Fable 5, the same model with new safeguards to block the most dangerous prompts from users. Three days later, Washington ordered it switched off.
David Sacks, the Silicon Valley venture capitalist and co-chair of Trump’s science advisory council, laid out the government’s case on X. Fable, he wrote, is “Mythos with guardrails,” and Anthropic had publicly promoted Mythos as a cyberweapon that should be regulated as one.
The two are essentially the same model; strip away Fable’s filters and you have Mythos. By Anthropic’s own account, Mythos can read software, find its security holes and exploit them faster than human experts. In the software trade, getting past those filters is called a jailbreak.
Amazon, a trusted partner of the government and the world’s leading “cloud” services provider online, found such a flaw in Fable, a way through the guardrails. Anthropic characterizes the jailbreak as narrow, and says the same trick works on its rivals’ models. “We disagree,” Anthropic said in a statement, “that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.”
But by the government’s logic a crack in the guardrail is a crack in a weapon, and the size is beside the point. “If there is a vulnerability – big or small – it is Anthropic’s responsibility to patch,” Sacks wrote. The government asked Anthropic to fix the flaw or take the model down, but Sacks said that Dario refused. Anthropic says it was never given specific flaws to fix and never refused to fix anything.
The US government already has legal controls over weapons. Genuinely dangerous technology, such as the advanced chips that can guide a missile, or the centrifuges that can enrich uranium for a reactor or a bomb, moves through a standing system: control lists drawn up in advance, licenses, published criteria, all of which a company can read and challenge. Anthropic had none of it.
The government’s letter cited the Export Control Reform Act, written in 2018 to bring Cold War protections into the 21st century and keep sensitive technology out of the hands of adversaries. The government has never turned that power on this kind of software model. The doctrine it leans on would treat letting a foreigner use controlled technology, even a person sitting at a desk in Boston, as the equivalent to shipping it to China or Russia.
But if letting a foreigner use an American service is an “export,” any service can be an export, and any company can be ordered to go dark.
A power this sweeping does not stop at borders. For decades, the American bargain on sensitive technology rested on a line between friend and foe. Adversaries were blocked; allies were waved through. But this order against Anthropic drew no line at all between the two. It treated a researcher in London, a banker in Tokyo, an engineer in Berlin and a foreign-born software developer in Seattle as a single forbidden class. America’s allies heard the message clearly. Kanishka Narayan, the British minister responsible for AI, whose own brief includes the safety institute that had spent weeks helping Anthropic test the model before it was pulled, was straightforward. “This week, the most advanced AI in the world was cut off from everyone in Britain,” he said. “Not by us, but by a decision taken in another country.”
The United States trusts these allies with its deadliest secrets, such as the technology behind its own nuclear submarines and fighter aircraft. Yet on a commercial chatbot they were handed the same blackout as China, an authoritarian rival.
Now set that against what the Trump administration is willing to sell China. It has cleared powerful Nvidia AI chips, a generation or two behind its best, for sale to Chinese buyers, and stands to take a quarter of the proceeds for the Treasury.
The logic is upside down. Washington is selling its adversary China chips it cannot reproduce on its own, and blacking out, for allies, the AI software it has no power to contain anyway. A chip is a physical object. It has to be fabricated, and China still cannot fabricate the best ones, which is precisely why a control on chips hurts China. An AI model is different. Strip away the hardware it runs on and what’s left is essentially a file – a very large set of numbers that can be copied perfectly and endlessly, at almost no cost. Once it leaks or a rival trains a cheaper version off its outputs, it spreads for free. China’s best AI models trail the American ones by months.
The strongest case against Anthropic came not from the government but from the media website Semafor. Citing a person familiar with the matter, its reporter Reed Albergotti wrote that the shutdown was driven partly by fear that a “China-linked” group had reached Mythos and might reverse-engineer it. The outlet did not specify which organization, how it gained access, or how the White House reached that conclusion.
Anthropic disputes the account. It says the government never raised Chinese access in the talks that produced this order, and that it already blocks its products inside China. Suppose the fear is accurate anyway. If a known hacker breaches a system, you go after that actor and seal up the way in. The government instead reached for a blackout on every non-American.
Switch off American AI overnight and you do not make the world safer. You teach allies to stop depending on you. They will build their own models in Europe and India, or turn to the open ones anyone can download, or, in time, switch to China’s AI models, which trail the American labs by a few months.
None of this means the hawks are wrong about the danger. Frontier AI is not ordinary software, and a model that can hunt down security flaws and string them into a cyber-attack may deserve real guardrails. The officials who lose sleep over it are being reasonable.
But acting fast and acting in the dark are not the same. A government that can switch off the world’s most advanced technology overnight, show no evidence for the decision, publish no findings, and answer to no one has joined the threat, not contained it.
Geoffrey Cain’s Steve Jobs in Exile: The Untold Story of NeXT and the Remaking of an American Visionary is out now.
Comments