Anthropic's most powerful AI lasted 76 hours. Then the US government shut it down

One of the world's best AI tools vanished in three days, killed by a single letter. How it happened is dramatic. Why it happened is something every business should understand.

Three days, then gone

On June 9, Anthropic released Fable 5, the smartest AI it had ever made available to the public. On June 12 at 5:21 PM, the US Commerce Secretary, Howard Lutnick, sent the company's CEO, Dario Amodei, a letter. By midnight, Fable 5 and a more advanced version called Mythos 5 were switched off, for every user in the world. Why the whole world, and not America? Because the government wanted to block foreign users, and that turned out to be impossible to do halfway. An AI like this works over the internet, where there is no passport check at the door. So an order meant for one country ended up shutting the tool down for the entire planet, including close US allies. Anthropic says it was given just 90 minutes to comply, with no real explanation of the danger.

A "jailbreak," or not

The official trigger was a "jailbreak." In simple terms, that means someone found a trick to make the AI ignore its own safety rules and do something it is supposed to refuse. Anthropic insists the trick was minor. It says someone asked the AI to look through some computer code and point out weak spots, something other popular AI tools, including OpenAI's, can already do anyway. The government disagreed. A senior White House adviser, David Sacks, said a trusted partner had found a serious jailbreak, that the government asked Amodei to either fix it or pull the tool, and that he refused. "That is not what the trusted partner and the US government believe," Sacks said, brushing aside Anthropic's "it's not a big deal" defence.

Then came the twist that changes the whole picture. That trusted partner was Amazon.

Why Amazon is the real problem

Amazon is not a neutral bystander here. It is one of Anthropic's biggest financial backers, having promised up to $4 billion. It also runs the computers that Anthropic's AI lives on. And now it is also the company that found the flaw and personally took it to the government, getting Anthropic's product shut down.

So Amazon is, all at once, an investor in Anthropic, its landlord, and the one who effectively got it regulated. There is not even a clean word for that kind of conflict yet, which shows how new and messy this whole industry still is.

The trap Anthropic set for itself

Here is the part worth pausing on, because it is the real lesson. For years, Anthropic built its brand on being the "safety-first" AI company. It kept describing its own technology as so powerful that it was almost a weapon, a threat to national security. That image helped it win cautious investors and big government contracts. But it also handed the government a ready-made reason to act. One researcher put it perfectly:

If you describe your product as a weapon in every press release, eventually a government takes you at your word.

And the timing was almost unbelievable. On June 10, just one day after launch, Amodei published an article asking the US government to give itself the power to block dangerous AI tools, the same way regulators ground unsafe aeroplanes. Two days later, the government used exactly that power, on him.

Trust was already breaking

The government was not even the first blow that week. Hidden in Fable 5's own fine print was an awkward detail: for certain questions, the AI would quietly give worse answers on purpose, without telling the user. It affected only a sliver of cases, but the idea outraged the developers who rely on it. An AI that secretly dumbs itself down, one expert said, is broken in its own way. Anthropic backtracked within two days, but trust was already shaken before the shutdown even arrived.

What this really teaches you

The ban will probably be lifted soon. The government has signalled it wants the tool back once the flaw is fixed.

The lesson is not to fear AI or avoid it. It is simpler: never let your work depend on one single AI. If everything you do only works with one company's tool, you are one letter away from losing it overnight. Build your process so you can switch tools easily, and the AI becomes what it always should have been, replaceable.