On June 12, 2026, the U.S. government invoked national-security export authorities to order Anthropic to suspend all access to its Fable 5 and Mythos 5 models, for every foreign national, anywhere, including the company’s own employees. Anthropic complied within hours, though they posted an auspicious notice for the world to view, creating some public disagreement. The case for the government export controls have always reached intangible technology and software, and a frontier AI model is exactly the kind of dual-use capability the regime exists to govern. If officials concluded that a jailbreak could unlock cyber capabilities to foreign persons, a fast, sweeping suspension is the conservative move: you stop the deemed export first and adjudicate the technical merits later. Under this reading, the directive is the system working as designed, under very broad authority, exercised quickly, to prevent diffusion of sensitive capability before harm occurs. The government’s case is persuasive.
Anthropic’s apparent counter is that the trigger was a narrow, nonuniversal jailbreak, essentially asking the model to find flaws in a codebase, producing capability already available from other deployed models, including competitors. It argues its defense-in-depth safeguards make Fable’s risks comparable to tools already on the market, and that recalling a product used by hundreds of millions over a minor finding sets a standard that would halt frontier deployment industry-wide. Its deeper objection is in fact more procedurally grounded: the company supports a government power to block unsafe deployments, but one exercised through a process that is transparent, fair, and grounded in technical facts. A verbal directive citing undisclosed concerns, it says, meets none of those tests.
Both positions accept the same premise, in that government may restrict dangerous AI exports, but they diverge on threshold and process. How serious must a vulnerability be to justify a recall? Who decides, on what evidence, and with what chance to respond? Export law has answered these questions for chips and encryption for decades. It has barely begun to answer them for models that update weekly and ship to the whole world at once.
When a single agency can disable a globally deployed AI model overnight on national-security grounds, where should the line sit between precaution and overreach, and what process would make those vested on both sides trust the decision either way?
Learn more about our work on international trade issues.
