Why the US Government Shut Down Claude and How to Prevent It

On Friday, June 12, 2026, at 5:21 in the evening, the United States Commerce Department sent Anthropic a letter. By the next morning two of the company's most capable models, Claude Fable 5 and Mythos 5, were gone. Not throttled, not restricted to certain users. Gone, for everyone, because that was the only way the company could comply with what the letter demanded. One directive, delivered to one company, took a frontier AI offline for every person on Earth who relied on it, in a matter of hours, with no court asked first.

The Letter

The order was an export control. It barred any foreign national, anywhere, from accessing Fable 5 or Mythos 5, including Anthropic's own foreign employees. There is no clean way to enforce a rule like that on a live service, so Anthropic disabled both models worldwide for all customers to be certain it complied. The government, by its own account, had a national security concern. It did not put the concern in the letter.

What Anthropic says it received instead was verbal: a claim that someone had found a way to bypass Fable 5's guardrails by asking the model to read a codebase and fix its flaws. Anthropic reviewed the demonstration and reported that the same capability exists in other widely available models, including OpenAI's GPT-5.5, and is used every day by the security engineers who keep systems running. The company called the recall a misunderstanding and made the structural objection out loud: pulling a commercial model used by hundreds of millions of people over a narrow, non-universal bypass, if it became the standard, would halt every new model release across the industry.

Set aside the question of whether the bypass was real. One fact survives it. A single government office, with a single letter to a single company inside its borders, switched off a frontier model for the entire world in an afternoon. No judge reviewed it first. The off switch existed because the model lives in one place, owned by one company, reachable by one directive.

The Pattern

A switch that powerful invites the question of whose hand is on it, and the record around Anthropic does not suggest a neutral one.

This was not the first move against the company. In February, after Anthropic refused to let its models be used for mass domestic surveillance or fully autonomous weapons, the administration ordered federal agencies to stop using its products and the Pentagon branded it a supply-chain risk, a label normally reserved for foreign adversaries. Within hours, OpenAI announced a deal to put its own models on classified military networks. When the dispute reached court, a federal judge found the government's actions "appeared to be designed to punish Anthropic" rather than to protect national security, called the supply-chain designation "Orwellian," and noted the Pentagon's own records tied it to the company's "hostile manner through the press." She blocked it. The pressure on Fable 5 began before the letter as well: a White House adviser said the administration had asked Anthropic to patch or pull the model the week before, and the company declined.

Some of the conflicts sit closer than the politics. The retired general who once led the National Security Agency, and who publicly doubted the case against Anthropic, also sits on OpenAI's board. As agencies dropped Claude, they added OpenAI and xAI in its place. And the recall arrived while Anthropic was preparing to go public, threatening the timing and the revenue of the offering. None of this proves the export order was written to wound a competitor or reward a friend. It establishes something narrower and harder to dismiss: the same office can reach the same switch on a thin report while standing to gain from the result. The hazard lives in the architecture, not in any one administration.

The Switch No One Owns

The defense against a captured off switch is not better hands on it. It is a system that does not have one.

When a model's weights are published openly, the way Meta's Llama, DeepSeek, Alibaba's Qwen, and Mistral publish theirs, the model stops being a service that a company can be ordered to suspend and becomes a file. Once that file is downloaded it sits on thousands of machines across every jurisdiction at once. There is no server to unplug and no company to compel, because there is no central copy left to revoke. This is not a fringe reading. The United States government's own telecommunications agency concluded that once weights are widely released they cannot be un-released, and noted that Mistral had distributed one of its models over BitTorrent, a protocol built specifically to evade control by any single party.

Decentralized AI is not a world where AI answers to no one. It is a world where no single office can switch it off for everyone on its own word. Reining in a genuinely dangerous model, or refusing to run it, becomes a matter of consensus among the many independent operators who host and serve it, instead of a letter to one executive. That is slower than a directive, and the slowness is the safeguard. A unilateral switch can be flipped on a misread report, or a conflicted interest, in a single afternoon. A distributed one requires that enough independent parties agree the danger is real, which is the bar a free society is supposed to clear before it takes a tool away from everyone at once. It is the same logic that makes trustless systems safer than trusted ones: remove the need to rely on any single party's good faith.

The capability the government said it feared was already everywhere, in GPT-5.5 and in open models anyone can download. Pulling one company's product did not remove that capability from the world. It only demonstrated how easily one company's product can be removed. Centralized AI concentrates not just the compute and the data but the control, into a point one letter can reach, and concentration is the fault line that runs through every centralized system. A model you can hold, running on hardware you own, answers to the people who run it and cannot be switched off by a letter. That is not AI beyond accountability. It is AI beyond the reach of a single hand.