Anthropic's most capable model, kept tightly restricted because of its skill at finding software bugs, was disabled globally on Friday through an export control order whose technical basis has not been put in writing.
On Friday at 5:21 pm ET, the U.S. government told Anthropic to cut off two of its most powerful AI models for every user, everywhere, and to do it immediately. The directive, reported by TechCrunch as disclosed by Anthropic on X, forced the company to disable Claude Fable 5 and Claude Mythos 5 worldwide, not just for the foreign nationals that export-control orders are normally aimed at. Anthropic complied. It also made clear, on X and in a blog post, that it believes the government got this one wrong.
The order landed because of what a government official described, verbally, as a "potential narrow, non-universal jailbreak" affecting the two models. A jailbreak in the AI context is a prompt that bypasses a model's safety guardrails. What that meant in practice, by Anthropic's account, was that the models could be coaxed into reading a codebase and identifying security flaws. That is a capability the company says is already widely available in tools and people.
The mechanism matters more than the headline. The legal vehicle was an export-control order, the kind of authority the U.S. uses to keep sensitive technology out of the hands of foreign adversaries. But the action it authorized was not targeted at foreigners. It was a global shutdown that pulled the models from tens of thousands of vetted U.S. users who were never the stated target. Other Anthropic models remain accessible.
The two models are not interchangeable. Claude Mythos 5, previewed in early April 2026 and kept tightly restricted, is Anthropic's most capable system. The company has said Mythos found flaws in every major operating system and web browser it was tested against, and shared it with roughly fifty vetted organizations under a program called Project Glasswing. Claude Fable 5 was released three days before the shutdown and is described as Mythos with additional guardrails. It had recently sat at the top of the Vals AI public benchmark, a widely watched leaderboard for AI coding assistants.
That is the asymmetry of the public record. The government has shared only a verbal description of a narrow technical flaw. Anthropic has put its own model lineage, its testing claims, and its compliance on the record. The technical basis for pulling two flagship products from a company's entire user base, on a Friday evening, is not in writing.
The action is also notable for the role Anthropic's own safety posture played in producing it. The company built Mythos with the explicit recognition that its ability to find software vulnerabilities was exceptional. It chose a tightly restricted release. The government's response, executed through an export-control mechanism rather than a public safety proceeding, treats the company's own risk assessment as the trigger for an immediate, global cutoff.
What the public is and is not being shown is now the question. A regulator that acts on a self-flagged AI safety risk, through an export-control order, and provides only a verbal summary of the technical basis, leaves the public with no way to weigh whether the action matched the risk. The next round of this fight will turn on whether that gap closes.