Last week's freeze on Anthropic's Fable 5 and Mythos 5 looked at first like a security scare with a short shelf life. A week on, the durable story is not the jailbreak that triggered it but the legal lever Washington pulled to act. In ordering Anthropic to obtain government approval before any foreign national could so much as use the two models, Commerce Secretary Howard Lutnick reached for export-control law and applied it somewhere it has never been applied: to the act of using a cloud-hosted program, rather than to the software code itself.
That distinction sounds like lawyerly hair-splitting. It is not. For years the technology industry operated on a settled assumption, backed by past official advisory opinions, that export controls covered the transfer of software but not the use of a cloud service running on someone else's servers. "You can't say that any more," Kate Koren of the Center for Strategic and International Studies told Bloomberg. The relevant doctrine, known as a "deemed export," has long forced companies to seek approval before sharing sensitive technology with foreign employees. Stretching it to cover a user typing into a chatbot is, as trade lawyers put it, uncharted territory.
Koren's verdict captures the unease. The government, she said, solved a real problem, the risk that Fable 5 could be jailbroken into a potent cyberweapon, but "opened a can of worms by doing it this way." Once Commerce has asserted the power once, no company can assume it will not reach for it again. "Any customer has to assume this could happen for any model at any time." The statutes Lutnick cited are real but largely untested, and a former Commerce adviser now at Morrison Foerster noted that the baseline reading of the law is still that cloud access is not an export at all.
The move is also a sharp reversal. Barely two weeks earlier, a Trump executive order had sketched a voluntary system in which labs would opt into having their models vetted. The Anthropic order replaced that with something compulsory and improvised, and the company complied within hours, disabling both models worldwide. Awkwardly, Anthropic does not verify users' citizenship and counts foreign nationals among its own staff, which gives a sense of how blunt the instrument is relative to the precision the theory demands.
Abroad, the gambit may already be self-defeating. The crackdown dominated talk at the G7 summit in France, where leaders renewed calls for technological sovereignty and floated routing future access through trusted partners. In Beijing the mood was closer to glee. After releasing the open-weight GLM-5.2, Zai founder Tang Jie predicted China would reach "Mythos-class" capability before the first quarter of 2027, even sooner than Elon Musk had wagered. Chinese labs are already pushing past a trillion parameters on domestic chips. A wall works only if there is something on the other side that cannot be built independently.
So the lasting question is less about Anthropic than about the precedent. If using a model is now legally a kind of export, the line between operating software and shipping a weapon sits wherever Commerce chooses to draw it, at least until a court is asked to weigh in. The administration plugged one leak. Whether the theory holds water is a test that has not yet been run.