On a Friday evening in June 2026, the most capable AI models on the planet went dark. Not because they failed. Not because a company changed its mind. Because the US Department of Commerce sent a letter, and a private company in San Francisco had forty-eight hours to comply with the most sweeping order of its kind ever issued against a software product.

The order was simple in form and staggering in scope: suspend all access to Claude Fable 5 and Claude Mythos 5, Anthropic’s newest and most powerful models, for any foreign national, anywhere in the world, including Anthropic’s own non-citizen employees. Anthropic said complying selectively was impossible. So it disabled the models for everyone. Customers in finance, healthcare, government, and critical infrastructure lost access overnight, with no warning and no transition plan.

Here is the part that should bother you more than the headline does. Researchers in the UK had been testing these models. British hospitals were piloting them. A UK politician who had just resigned his post over a defense spending dispute put it bluntly: this week, the most advanced AI model on the planet got switched off by a foreign government, and the people who had been relying on it had no say in the decision at all. The European Commission opened an inquiry into the implications. Politicians across France and the Netherlands started calling for sovereign AI investment, framing the episode as proof that depending on someone else’s infrastructure means depending on someone else’s politics.

None of those countries got a vote. None of them got a phone call. A regulatory action taken inside one jurisdiction, for reasons that jurisdiction has still not made public, became instant foreign policy for every government, hospital, and company on earth that had built something on top of these models. That is the part worth sitting with. We talk about the internet, and now AI, as though they are global commons. They are not. They are licensed. They run on infrastructure owned by one country’s companies, regulated by one country’s agencies, and subject to one country’s politics, dressed up in the language of universality. The moment that country decides otherwise, the rest of the world finds out exactly how much sovereignty it actually had, which is none.

What makes this particular case worse is how thin the stated justification is. The government’s letter reportedly gave Anthropic only verbal notice of a “narrow, non-universal jailbreak,” with no written technical basis and no public evidence. Anthropic pushed back hard, arguing the underlying capability was already available in other public models with no export restrictions attached, and that the process behind the order had none of the transparency or evidence-based rigor that should accompany a decision with this much global reach. A cybersecurity researcher who reviewed the actual paper behind the claim later said the “jailbreak” amounted to asking the model to review code for security flaws, the kind of task defenders do every day, and that pulling the model over it actively makes cyber defense worse, not better. Reporting also tied the original tip to a security researcher connected to one of Anthropic’s own major investors, which tells you something about how tangled commercial interest, political grudge, and “national security” can get before anyone calls it what it is.

This is not the first round between Anthropic and this administration. Months earlier, the Pentagon had already labeled the company a supply chain risk, a designation usually reserved for foreign adversaries, after Anthropic refused to drop safeguards against its models being used for autonomous weapons or mass domestic surveillance. That fight is still in court. So when the export control order landed, it landed inside an existing pattern, not in a vacuum. Calling this a clean, neutral safety decision asks me to ignore the relationship it grew out of.

This is the question I actually want to sit with, more than the technical or legal one: how do we draw a line between a government protecting its citizens and a government protecting its leverage, when both get dressed in the exact same vocabulary? Separation of powers assumes the people exercising power are accountable to something outside themselves. What this week showed is that a handful of individuals, answering to no external technical standard and no international body, can make a decision on a Friday afternoon that reorders the operating conditions of hospitals and researchers on another continent who never had a seat at the table.

I keep coming back to the same uncomfortable place. It is not the technology that is untrustworthy here. It is the people running the systems around it, the same as it has always been. Self-interest dressed as principle, fear dressed as caution, leverage dressed as safety. The infrastructure changes. The thing underneath it, the wounded, self-protective, status-seeking core of human decision-making, does not. We built something unprecedented in Mythos and Fable, and we still managed to make the story about it depressingly familiar: a few people in a room, deciding for everyone else, while calling it protection.

If the smartest model on earth can be switched off by one office, in one country, on one Friday, then sovereignty was never the right word for what any of us had. We were borrowing power and calling it ours. The real question isn’t whether this happens again. It’s whether we ever stop pretending it couldn’t.