Amazon Got Anthropic's Fable 5 Switched Off

For a few days, everyone told the same story: the U.S. government forced Anthropic to switch off Fable 5, its most powerful model, in the name of national security. That part is true. But Washington's order is only the last line of a story that begins somewhere else, inside the house itself. The trigger was Amazon. Anthropic's first backer.

Let's take it from the top, because every single day matters here.

Tuesday, June 9: the launch of the year

Anthropic ships Fable 5, the public version of its Mythos model. The company frames it as its most capable system yet: the first to clear 90% on the Hex benchmark for long-form analysis, strong at code and vision. The pricing radiates confidence at 10 dollars per million input tokens and 50 on output, double Opus 4.8.

Above all, Anthropic sells safety as the headline. For months, the company has insisted that Mythos is too dangerous to circulate freely. So Fable 5 ships wrapped in guardrails: on cyber, bio and chemical topics, the model refuses and falls back to Opus 4.8. Mandatory 30-day traffic retention, and one promise hammered everywhere, "no universal jailbreak across more than 1,000 hours of testing."

Tuesday is the show of strength. Nobody suspects that the very same promise is about to become the fuse.

Wednesday, June 10: the first crack

Twenty-four hours later, a Fable 5 jailbreak goes up publicly on X, posted by an account that specializes in this kind of bypass. The method: ask the model to read specific codebases and flag software vulnerabilities. Narrow, technical, a long way from a weapon of mass destruction. But the damage is done: proof that the "1,000 hours of testing" vault opens is now in plain sight.

This is where the character nobody expected walks in. It isn't the NSA or any intelligence agency raising the alarm. According to Fortune, it's researchers at Amazon. They reproduce a partial bypass of the guardrails, and Andy Jassy, Amazon's CEO, escalates the concern straight to the U.S. administration.

Sit with the irony for a second. Amazon has poured several billion dollars into Anthropic. It's the investor most exposed to the company's value, months out from an IPO. And it's the one laying the first stone of the machinery that's about to switch off the flagship product of its own stake.

Friday, June 12, 5:21 p.m.: the kill switch

The letter lands on a Friday, late afternoon. Department of Commerce, under export-control authority. The order bans access to Fable 5 and Mythos 5 for "any foreign national, whether inside or outside the United States," down to Anthropic's own foreign employees.

Telling a U.S. customer apart from a foreign one in real time across a global API simply isn't workable. So Anthropic has one option: cut everything, for everyone. Claude apps, the API, the GitHub Copilot integration. The company has roughly ninety minutes to comply, with no warning.

The backroom plays out over the phone. On one call, Treasury Secretary Scott Bessent reportedly warned Dario Amodei he was "making a bad decision." David Sacks, the White House AI adviser, coordinates the response. Three days after the launch of the year, Anthropic's most powerful model no longer exists for anyone on the planet.

Anthropic's argument holds, and still doesn't save it

On the merits, the company isn't wrong. Perfect jailbreak resistance exists on no model. The flaw is narrow, not universal. And reading code to find bugs is a capability available everywhere, starting with OpenAI's GPT-5.5, used every day by cybersecurity professionals. Anthropic warns that, applied to the whole industry, this standard would freeze the deployment of every new frontier model.

That's defensible. It doesn't erase the staging that made the move possible. You don't spend months selling the idea that your model is a weapon apart, then act surprised when a government treats it like one.

The model isn't flagged for permanent removal, and on June 16 Anthropic sent engineers to Washington to negotiate in person with the Department of Commerce. As these lines are written, access still hasn't been restored. The hangover, on the other hand, is already here.

What the episode shifts, everywhere

The most interesting part isn't the jailbreak. It's what the world watched in real time. EU countries and the United Kingdom were, in fact, talking to Anthropic about broader access to Mythos to harden their own systems. They learned that a tool framed as critical could be switched off by a foreign government in ninety minutes, on a Friday night.

The lines moved at once. The European Commission said it was reviewing the directive, its spokesperson noting that emergency measures "should not be discriminatory toward partners." Finnish MEP Aura Salla was blunter: Europe cannot build its technical potential "by relying on access that a foreign government can cut off overnight." In Canada, on the eve of the G7, Mark Carney flagged the danger of "over-reliance on a limited number of U.S. suppliers."

None of these reactions are about the safety of a model. They're about the architecture of power the episode just exposed. A growing share of the world's cognitive infrastructure sits with a handful of U.S. players, inside a jurisdiction that can decide, on a Friday at 5:21 p.m., that part of the planet no longer has access. The Fable 5 jailbreak will be patched or routed around. The switch stays in Washington.