On June 12, 2026, the US Commerce Department sent a letter to Anthropic ordering it to immediately block all foreign nationals from accessing Claude Fable 5 and Mythos 5, its two newest and most capable AI models. Anthropic had no practical way to check the citizenship of every user in real time. So it shut both models down for everyone, everywhere, that same evening.
The shutdown lasted 19 days. It ended on June 30 after Anthropic strengthened the safety controls on the models. Access to Fable 5 was restored globally on July 1. The more restricted Mythos 5 is now available only to approved US organizations.
The trigger was a reported security vulnerability: a technique that could bypass the models' built-in safety limits, allegedly discovered by Amazon researchers and flagged by Amazon's CEO to federal officials. The government said Anthropic refused to fix it. Anthropic disputed that characterization entirely, saying the flaw was narrow and that it was never given specific details before the order arrived.
For enterprises that had already built operational workflows on Fable 5, which had only launched three days before the shutdown, access vanished with no warning, no compensation, and no clear timeline for return. Finance, healthcare, and critical infrastructure customers were among those cut off.
The legal question hanging over all of this is whether the government actually had the authority to do what it did. The law it used, the Export Control Reform Act of 2018, was written to control physical exports like weapons parts and restricted hardware. Past Commerce Department guidance had treated remote access to software running on US servers as outside the reach of those same controls. Legal analysts at Mayer Brown noted that extending export control jurisdiction to AI model access is entirely novel territory, with the legality far from settled.
The procedure used was also unusual. The standard mechanism is normally applied to specific transactions with specific countries, not to every foreign national on earth simultaneously. The order was secret, had no expiration date, required no sign-off from the Defense or State departments, and was not published anywhere. All of that stands in contrast to the normal, slower process that includes public comment and coordination with US allies. European officials and others expressed alarm during the blackout about their dependence on Washington's policy decisions for access to the world's most capable AI tools.
Anthropicchose to negotiate rather than sue. The company had previously sued the Defense Department over a separate dispute, so the negotiation was likely strategic. Anthropic's CEO had also published an essay just days before the shutdown arguing that governments should have the power to block AI models deemed too dangerous, which made a legal challenge to that same principle awkward.
The broader backdrop is a US government still figuring out what AI oversight should look like. President Trump signed an executive order on June 2, just ten days before the Anthropic shutdown, creating a voluntary framework under which AI developers could give the government 30 days of early access to powerful new models before public launch. The order explicitly says this is not a mandatory licensing or approval system. But the Anthropic episode shows that the government does not necessarily need a formal licensing law to act as a gatekeeper: export control authority, applied through a private letter, achieved the same result.
The practical lesson for any organization that runs important work through top-tier AI tools is straightforward: the access you pay for is not guaranteed. It can be cut off by a government order that has nothing to do with you, your contract, or the quality of the tool. The 19-day Anthropic shutdown is now a documented example of that risk becoming real. Organizations that had backup options or could fall back to earlier model versions weathered it. Those that had built tight dependencies on Fable 5 specifically had no good options.
The Chinese angle is also worth noting. The shutdown handed three weeks of competitive breathing room to Chinese AI developers who are closing the capability gap with US models. The restriction that was meant to protect US interests may have quietly helped the competition it was designed to contain.