The U.S. Government’s Sudden Clamp‑Down on Anthropic’s Biggest AI Models

When a terse letter from the U.S. Commerce Department hit Anthropic’s inbox on June 12, 2026, the company had almost no breathing room. The memo, citing national‑security worries, told the AI start‑up to stop every foreign national – even its own non‑citizen engineers – from accessing Claude Fable 5 and Mythos 5. Within a few hours the two models went dark for the entire world.

Anthropic’s options were painfully limited. The department’s demand was clear: you can’t keep a running list of who is a foreign citizen and who isn’t, especially when users may be logging in from a coffee shop in Bangkok or a dorm room in Berlin. So the safer (or perhaps the only) move was to pull the plug for everyone.

The shutdown lasted until June 30, when the Commerce Department finally lifted the ban – but only after Anthropic had added a raft of safety guardrails that, frankly, crippled the models. The same tests that once showed Claude Fable 5 breezing through complex reasoning now saw it stumble on three out of twelve routine tasks. In short, the AI’s “brain” was throttled so hard that even harmless questions tripped the new defenses.

This episode sent a shiver through the U.S. AI sector. As a scholar who studies technology law, I keep asking two kettle‑of‑fish questions: first, does the U.S. government even have the authority to act as a global gatekeeper for AI? And second, was the June 12 order itself on solid legal ground?

The letter was issued under the Export Control Reform Act of 2018 – a law originally written to keep dangerous hardware like uranium‑enrichment centrifuges out of the wrong hands. The act also gives the Bureau of Industry and Security (BIS) power to require export licences for “emerging and foundational technologies.” In this case, BIS treated access to Anthropic’s cloud‑based models as a kind of export.

That raises a knotty issue: what exactly is being exported? When you type a prompt into Claude, the model stays on Anthropic’s servers; all you receive is a text reply. The model itself never leaves U.S. soil. Past BIS guidance has usually said that remote access to software hosted domestically falls outside export‑control jurisdiction. The fact that Congress is now debating whether to change that suggests the current law might be outdated, but it doesn’t automatically give the government a clean sweep.

Even if we accept that a model’s output counts as an export, the way the order was applied looks… over‑broad. The so‑called “is‑informed” mechanism traditionally notifies a company that a specific transaction to a particular country needs a licence. Here, the directive swept every foreign national on the planet under one blanket – an approach that may exceed the authority granted to BIS.

Historically, the U.S. has moved more deliberately when it comes to restricting dangerous tech. Multiple agencies would weigh in, public comment periods would be opened, and any measures would be published in the Federal Register. In emergencies, the government can invoke a temporary classification (code 0Y521) that still requires sign‑off from the Defense and State departments and expires after a year unless renewed. The Anthropic letter, by contrast, was unilateral, secret, and open‑ended – a far cry from the usual checks and balances.

What’s perhaps even more surprising is that Anthropic didn’t fight the order in court. The company called the episode “a misunderstanding” and sent its senior team straight to Washington for negotiations. That may sound like capitulation, but Anthropic has long echoed the government’s own concerns. Just two days before the notice, CEO Dario Amodei published an essay urging regulators to have the power to block or deter the deployment of “frontier” AI models deemed too risky. Challenging the order would have undercut that very stance.

There’s also a practical snag: the 2018 Export Control Reform Act strips federal courts of the usual power to strike down such export‑control decisions as arbitrary. Any challenger would have to prove the order was not just unreasonable, but plainly unlawful – a high bar that discourages litigation.

All told, the Anthropic saga marks a turning point. Whether the Commerce Department’s move was fully legal or not, it shows a willingness to use export‑control tools to police AI in real time, even at the cost of stifling innovation. As AI systems become more capable and more embedded in daily life, we can expect the debate over who gets to decide who can use them – and on what terms – to heat up dramatically.