Anthropic disabled Claude Fable 5 and the related Mythos 5 globally, including for US customers, after a national security letter. The company says the cited vulnerability is minor and already public.
Late Friday afternoon, a US government letter ordered Anthropic to cut off access to two of its AI models: Claude Fable 5, a large language model released the previous Tuesday with built-in safeguards blocking questions about cybersecurity, biology, and chemistry; and a related model called Mythos 5. The order cited national security concerns. Anthropic complied globally, then within hours published a blog post disputing the technical premise of the action.
The dispute centers on a "jailbreak" — a technique for getting a model to ignore its built-in safety filters. According to Anthropic's account, the jailbreak at issue surfaces only minor, already-documented vulnerabilities that other publicly available models can also find. The company argues the order overreaches both in targeting that finding and in scope: the directive as conveyed asked Anthropic to suspend access for "any foreign national, whether inside or outside the United States, including foreign national Anthropic employees." To be safe, the company removed access for all customers, not just foreign nationals. That global cut, including US users of a US-built product, is what Anthropic calls unprecedented.
The government has not publicly detailed the specific concern that triggered the letter. The action lands on a company already in conflict with the Trump administration. Earlier in 2026, the Department of Defense designated Anthropic a "supply-chain risk" after the company tried to draw red lines on US military use of its technology, a designation that effectively bars government agencies and contractors from working with the lab. Anthropic responded with lawsuits against the administration, the same set of disputes that frame Friday's letter as another flashpoint.
What makes this more than a company-versus-administration story is the category of power being used. The Friday letter functions like an export-control-style directive, the same legal pathway that lets the Commerce Department restrict chip sales to specific countries, but pointed inward at a domestic lab and a publicly released product. The statutory basis, the issuing agency, and the review path are not yet clear in public records. WIRED's reporting, which Anthropic has not disputed on the underlying facts, characterizes the order as "unprecedented" in scope for this kind of letter.
The open question, the one that matters to anyone building or using a frontier model, is what evidentiary standard now applies. The government has asserted a national-security concern and has not specified it. Anthropic says the cited vulnerability is technically weak. If the trigger is a publicly reproducible jailbreak on a model already shipping with content restrictions, the precedent extends to any future release. If the trigger is something the government can describe in a sealed filing, the precedent still sets a default, with the burden of disclosure shifted to the company rather than the state. The next model release from any frontier lab, and any appeal Anthropic files, will test which frame holds.