The overnight off-switch: Washington pulls two Anthropic models, and reaches inside the lab to its own foreign staff

The sequence is worth stating plainly because the speed is part of the story. Anthropic launched Fable 5 and Mythos 5 on Tuesday as the most capable systems it had ever shipped — Fable 5 to the general public behind heavy safeguards that fall back to a weaker model in cybersecurity and biology, Mythos 5 without those guardrails to a small set of trusted cybersecurity and infrastructure partners. Three days later, at 5:21pm on a Friday, a letter arrived from the Commerce Secretary, drafted with the Bureau of Industry and Security, citing national-security authorities it did not specify, ordering the company to deny the models to every foreign national on earth, its own staff included. By Friday night both models were dark for all users. There was no rulemaking, no comment period, no published classification decision — an export-control directive functioning as an emergency kill switch for a product already in the hands of hundreds of millions of people. For any government that builds or buys frontier AI, that capability is the headline, independent of whether this particular invocation was justified.

The proximate trigger, as best the public record allows, was a single disputed jailbreak. Anthropic's account is that the government believed it had learned of a way to bypass Fable 5's safeguards, and that the demonstration it was shown consisted of asking the model to read a specific codebase and fix the software flaws in it — a narrow, non-universal technique that surfaced only previously known, minor vulnerabilities, of a kind other public models including OpenAI's GPT-5.5 will also produce and that defenders use every day. A separate, louder claim came from the jailbreaker known as Pliny the Liberator, who said he had “liberated” Fable 5; Anthropic disputes that this is a real jailbreak, arguing the core safeguards are enforced by independent classifiers that survive the model's conversational refusals being coaxed away. Whether or not the government's basis and Pliny's claim are the same artefact, the salient point for the trade-policy reader is the standard being applied: a recalled commercial model, industry-wide, on the strength of a contested narrow finding. Anthropic's warning that this would “essentially halt all new model deployments for all frontier model providers” is not special pleading; it is a description of where the precedent points.

The mechanism that makes the order bite inside the company is the deemed-export rule, and this is the part that turns an export story into a labor-and-immigration story. Under the EAR, releasing controlled “technology” or source code to a foreign national located in the United States is treated (“deemed”) as an export to that person's country of nationality. Nothing has to cross a border. A screen share, an oral briefing, a granted permission on a repository, a glance at a model endpoint: each is a release, and each release to a non-US person is an export. The only people exempt are US citizens, lawful permanent residents and a narrow category of protected individuals. Everyone else (the H-1B research scientist, the O-1 hire, the recent PhD on OPT, the engineer on an L-1 transfer) is, for the purposes of this rule, a foreign destination that happens to sit at the next desk. The directive's own language, “including foreign national Anthropic employees,” is the doctrine stated out loud. And because a public API cannot reliably verify the nationality of the person behind a key, the only compliant response to a deemed-export-style restriction on a mass-market model is to turn the model off for all, which is exactly what happened. A nationality restriction on a cloud product is, in practice, a total recall.

That is why the talent arithmetic is not a footnote. By MacroPolo's tracking, around two-thirds of the elite AI researchers working in the United States completed their undergraduate education outside it, and the largest single source country is China — the country at the most restrictive end of every tier Washington has ever drawn. A nationality-keyed control regime does not graze the edge of a frontier lab's workforce; it runs through the center of it. The structural response is the one cleared defense contractors have used for generations, now imported into an industry built on the opposite premise: segregated access, US-person-only project rooms, walled compute, technology control plans, and an HR-and-legal apparatus that screens by citizenship before granting access to the crown-jewel model. An industry whose advantage was assembling the best researchers on earth regardless of origin now has a standing regulatory incentive to sort them by origin, and a live demonstration that the government can force the sort overnight.

The second-order effects radiate outward, and several cut against Washington's own aims. A model that can be switched off by letter is a different commercial proposition for an enterprise buyer, who now has to price regulatory availability alongside uptime; expect procurement teams to ask frontier vendors what their continuity plan is when the next directive lands. The deemed-export logic gives the most mobile foreign talent a sharper reason to take the job in London, Paris, Toronto or Abu Dhabi rather than wait out a license — a quiet reverse brain drain dressed as compliance. The discretion on display, an unspecified authority invoked on a Friday night against a company already suing the administration, will not be lost on allied governments weighing how much of their AI stack to anchor in a single jurisdiction; it strengthens every sovereign-AI and “trusted compute” pitch already in motion. And there is a chilling edge for the open security research that Anthropic's own safeguards depend on: if demonstrating a narrow jailbreak can trigger a national-security recall, the incentives around disclosure get more complicated for everyone.

For the practitioner community Doyen serves, the reframing is clean even if the politics are not. Frontier AI has just been shown to be an export-controlled, politically contingent industry, and it should be modelled like aerospace, nuclear and advanced semiconductors rather than like software-as-a-service. That changes the locational calculus in concrete ways. Nationality becomes a workforce variable that AI investors must weigh when they site research, advantaging jurisdictions with deep US-person-equivalent talent pools or trusted-partner arrangements and penalising those whose pitch was cheap global talent with no security overlay. Jurisdiction itself becomes a product feature: “trusted cloud” status, allied-tier designation and a credible continuity story are now things a serious AI-investment proposition has to offer, ahead of the tax abatement. And the open question every economic-security team should be asking is procedural — under what authority, on what evidentiary standard, and with what notice can a deployed model be pulled — because Friday's answer was, in effect, “a letter, a contested finding, and none.”