Two frontier AI models went dark worldwide on Friday. The order that did it is not public.

On Friday, June 12, at 5:21 p.m. Eastern, Anthropic received an instruction from the United States government. By the end of the night the company had switched off its two most capable artificial-intelligence models — Claude Fable 5 and Claude Mythos 5 — for every customer it has, in every country it serves. Not throttled. Not gated behind a higher tier. Off. The models had been public for three days.

That is the part of the story that can be stated without dispute, because Anthropic put it in writing. In a short notice posted to its own newsroom over the weekend, the company said it had been served an export-control directive citing national-security authorities, and that the directive required it to suspend access to both models by any foreign national — inside or outside the United States, including the company's own foreign-national employees. Anthropic complied the same evening. Everything else about the order, including the document itself and the evidence behind it, is not something the public has been allowed to see.

I cover surveillance and the data economy by reading the paper that power leaves behind: the contract, the filing, the procurement record, the line in a budget no one expected to be read. This story is an unusual one for that method, because the most consequential act in it — a government reaching into a commercial product and turning it off for the entire world — has so far produced almost no paper at all.

What the order actually does

The mechanism matters, so start there. The United States did not pass a law on Friday, hold a hearing, or publish a rule. According to Anthropic's account, it invoked export-control authority — the same body of law that decides which advanced chips may be sold to which countries — and applied it not to a chip but to access to a model.

Export-control law has a concept built for exactly this kind of reach. It is called a deemed export: under the Export Administration Regulations, releasing controlled technology to a foreign national counts as an export to that person's home country, even if the release happens inside the United States, even if it happens over a screen. The rule was written decades ago for blueprints and source code handed to a visiting engineer. Applied to a frontier model, it means that letting a foreign national use Fable 5 — a German researcher in Berlin, a contractor in Bangalore, an Anthropic engineer on a visa in San Francisco — is, in the government's framing, an export of a controlled item to Germany, India, wherever the person is from.

That framing is what made a global shutdown the only available form of compliance. A company can wall a product off by country. It cannot reliably wall it off by the nationality of the person at the keyboard, because nationality is not something an API can see. Anthropic employs foreign nationals. Its customers employ foreign nationals. Once the order reached every foreign national everywhere, including the company's own staff, there was no surgical way to obey it. The off switch was the compliant switch.

A company can wall a product off by country. It cannot wall it off by the nationality of the person at the keyboard. Once the order reached every foreign national everywhere, the off switch was the compliant switch.

This is worth sitting with, because it is the actual lever. The headlines say the government banned two AI models. What the government did, more precisely, was reclassify access to those models as a controlled export and then bar the export to a category of person so broad that the product became impossible to keep running for anyone. The blast radius was not a side effect. It is how the instrument works.

The trigger was a jailbreak posted to X

The proximate cause, by every account including Anthropic's, was a single post. On June 10, an account that goes by Pliny the Liberator — a well-known figure in the loose community that publishes prompts to bypass AI safety guardrails — posted a jailbreak of Fable 5 to X. Two days later the directive arrived.

Here the record gets thin in a way that should bother anyone who cares about how decisions like this get made. Anthropic says the government's evidence of the danger was delivered verbally — that it has been shown what it characterizes as a 'potential narrow, non-universal jailbreak,' but not a written technical finding it can examine and answer. The company's description of the underlying capability is almost anticlimactic: in its understanding, the jailbreak amounts to asking the model to read a specific codebase and fix the software flaws it finds. Anthropic says the vulnerabilities surfaced were minor and already known, the kind of thing a defender does on an ordinary Tuesday, and that the same work can be done with other commercial models, naming OpenAI's GPT-5.5 specifically.

I want to be careful about what I am and am not asserting. I have not seen the government's evidence. Neither, by its own account, has Anthropic seen it in any form it can rebut on the merits. That asymmetry is the point. A company knows when its own product has been taken offline; it does not necessarily know why, in any form it could take to a court or a standards body. And the public knows least of all.

A weekend timeline

• June 9: Anthropic releases Claude Fable 5 and Claude Mythos 5, its two most capable models. Mythos is restricted from launch to vetted organizations.
• June 10: An account known as Pliny the Liberator posts a jailbreak of Fable 5 to X.
• June 12, 5:21 p.m. ET: Anthropic receives a U.S. government export-control directive citing national-security authorities, barring access by any foreign national anywhere.
• June 12, evening: Anthropic disables both models for all customers worldwide, calling it the only way to comply.
• Over the weekend: Anthropic publishes a statement disputing the severity, says it will share more within 24 hours and restore access 'as soon as possible.' No public order, no statutory citation, no written technical finding has been released.

Why the absence of a document is the story

In the kind of reporting I do, the question is always the same: show me where it says that. A wiretap has a warrant. A data-broker sale has a contract. An export ban on a chip has a rule published in the Federal Register, with a control classification you can look up and a comment process that left a record. Those records are often boring, sometimes redacted, occasionally fought over for years. But they exist, and their existence is what makes the power reviewable.

What is described here has, so far, none of that. There is a directive Anthropic has not published, evidence Anthropic says was spoken rather than written, and a legal theory — access as deemed export — that has not been tested in the open. The result was the most far-reaching action a government has yet taken against a commercial AI product: not a fine, not a recall request, not a mandated safety patch, but a worldwide off switch, executed inside an evening, on the strength of a finding no outside party can examine.

Not a fine, not a recall request, not a mandated safety patch, but a worldwide off switch, executed inside an evening, on the strength of a finding no outside party can examine.

Anthropic, for its part, is not framing this as a fight. Its statement is studiously cooperative — it believes the matter is a misunderstanding, it complied without litigating, it says it wants to restore access quickly and work with the government. But it made one argument that reaches past its own situation. If the standard applied to Fable 5 — that a model which can find and fix software bugs is dangerous enough to pull — were applied across the industry, the company argued, it would essentially halt all new model deployments, because every capable model can do that, and because, in its words, perfect jailbreak resistance is not currently possible for any provider. It asked that oversight of this kind run through transparent, fair, and clear statutory processes grounded in technical facts.

Translate that out of corporate register and it is a request for a paper trail. Not a softer outcome — a documented one. A rule you can read, evidence you can contest, a process that leaves something behind.

Who decides, and on what record

There is a real national-security question underneath this, and I am not going to pretend there isn't. A model capable enough to autonomously find and exploit software vulnerabilities is genuinely dual-use, and a government has a legitimate interest in who can wield one. Anthropic itself has spent the past year building that case — its own Project Glasswing program scans critical infrastructure for flaws, and the company has argued loudly that frontier models are reaching capabilities that warrant control. It is not in a strong position to be shocked that someone took the argument seriously.

But 'there is a real question' and 'this is how you answer it' are different claims. The thing that should worry you is not that the government acted. It is the shape of the action: a category of controlled item invented on the fly, defined so broadly it forced a global blackout; evidence held verbally; no published order; a turnaround measured in hours. Every one of those choices reduced the amount of record left behind. Each is individually defensible in an emergency. Together they describe a power that can switch off a widely used product worldwide and owe the public no document explaining why.

The models will almost certainly come back; Anthropic expects them to, and the dispute reads like one that gets resolved in a conference room. That is not the part to watch. The precedent is the part to watch. Friday established that access to a commercial AI model can be treated as a controlled export, that the control can be aimed at a class of person broad enough to take the product down everywhere, and that it can be done on a spoken finding. Whatever you think of the merits in this case, that capability now exists, demonstrated and unrebutted, and the next time it is used the target may have less standing and the cause may be less benign.

For a beat that runs on documents, the lesson of the weekend is the document that isn't there. The most powerful thing that happened to artificial intelligence in June was a model going dark on a Friday night, and the order that did it is something the rest of us are, for now, simply told about. Power leaves a paper trail. The story is what it means when this much power leaves so little of one.