America's first AI shutdown: inside the 19 days Fable 5 went dark

In 1954 the United States revoked the security clearance of the physicist who built its bomb; in 2026 it switched off the computer scientist's model, and the parallel is closer than either government would like.

In the spring of 1954, the Atomic Energy Commission spent four weeks deciding whether J. Robert Oppenheimer, the physicist who gave America the bomb, could still be trusted with its secrets. The hearing was about power more than security: the state, not the laboratory, decides what happens to a dangerous invention. The clearance was revoked.

Seventy-two years later, on the evening of Friday 12 June, the US Commerce Department ordered Anthropic to suspend access to Claude Fable 5 and Claude Mythos 5 for every foreign national on Earth, including its own non-US employees.verifiedVerified Sourced from Statement on the US directive to suspend Fable 5 and Mythos 5 access (Anthropic, 12 June 2026). Nationality cannot be checked in real time, so compliance meant switching both models off for everyone.verifiedVerified Sourced from Mythos AI broke into almost all NSA classified systems in hours (Security Affairs / The Economist, 22 June 2026). They stayed dark until 1 July. Nineteen days.

The episode ends, at least formally, on Tuesday, when the half-usage window that came with Fable 5's restoration expires and the model moves to metered credits. The three weeks in between revealed who actually controls AI.

The essay that aged in a day

On 10 June, one day after launching Fable 5, Amodei published Policy on the AI Exponential, arguing frontier models should face mandatory testing, with release 'blocked or reversed as a threat to public safety' when they fall short.verifiedVerified Sourced from Policy on the AI Exponential (Dario Amodei, 10 June 2026). It was a long-held position: his January 2025 DeepSeek essay cast export controls as among America's most powerful tools against China.

Two days later the power was used for the first time, on him. The regime he sketched assumed a regulator that was deliberate, technical and slow. The one that arrived moved in an evening and did not distinguish between his safety-hardened consumer model and the restricted variant beneath it.

Three words: fix this code

The trigger was almost comically small. Researchers at Amazon, Anthropic's largest investor, gave Fable 5 code with known vulnerabilities and asked it to 'review the code for security issues'. It refused, as designed. They rephrased: 'fix this code.' It complied, producing patches and, in one case, code showing how a flaw could be exploited.

Amazon chief executive Andy Jassy raised the finding on a White House call, officials steered him to Treasury Secretary Scott Bessent, and the administration demanded Anthropic fix the jailbreak or pull the model. Amodei declined, disputing that a narrow, already-known technique justified recalling a model used by hundreds of millions. The irony: the report that grounded Anthropic's flagship came from the company that had agreed weeks earlier to put up to another US$25 billion into it.

Security researchers largely sided with Anthropic: Katie Moussouris said the technique 'is not a guardrail bypass' but 'the most valuable thing an AI model can do for defensive security', and an Alex Stamos open letter drew about a hundred signatures from researchers at Nvidia, Adobe, Google and Sophos.

There is a competing explanation. Senator Mark Warner has said General Joshua Rudd, head of the NSA and Cyber Command, told him on 11 June that Mythos had broken into almost all of the agency's classified systems in hours during an authorised red-team exercise. No agency has confirmed it, and a self-test that finds flaws is not an intrusion. But if true, Washington was not reacting to a three-word jailbreak alone.

Nineteen days of dependency, made visible

Mythos 5 has been restored only for vetted US organisations, leaving Australian organisations locked out. The Australian Signals Directorate said it is 'working with US and other international partners on the implications of frontier AI models', and ACU's Niusha Shafiabady put it plainly: 'Access to overseas AI models is important, but it is not the same as AI sovereignty'.

The models were not withdrawn by their maker for safety reasons; they were withdrawn because one government decided the world's access was a bargaining chip. Every enterprise on a frontier model now knows its continuity turns on the temperature of one relationship with the White House.

A truce with conditions

The resolution came in stages: Mythos 5 returned on 26 June for more than a hundred vetted US organisations under the Glasswing program, then Commerce Secretary Howard Lutnick lifted the controls entirely on 30 June, saying his department had 'worked closely with Anthropic to analyze and approve Fable 5'. Anthropic agreed to pre-release future frontier models to federal reviewers, share vulnerability information, and build an industry safety framework with Amazon, Google and Microsoft.

It is, almost line for line, the regime the 10 June essay asked for. Amodei wanted an FAA for AI; what he got is the beginnings of one, negotiated at export-control gunpoint. The essay's argument won. Its author lost nineteen days of revenue and a measure of sovereignty to make the point.

Tuesday's cliff is about capacity, not punishment

Since 1 July, Pro, Max, Team and select Enterprise subscribers have had Fable 5 at up to half their weekly limits. That window closes on Tuesday 7 July; from Wednesday the model runs on usage credits on those plans, while staying fully available on the API and consumption-based Enterprise plans.

Anthropic insists this is capacity management, not stealth withdrawal, and aims to restore standard subscription access once capacity allows. But the make-good is thinner than the promise: at launch, subscribers were told Fable 5 was included through 22 June, the shutdown cut that to roughly three days, and the replacement is seven days at half usage. Capacity explains the cliff; it does not restore what was taken.

For Australian teams the advice is unglamorous: use the window today and tomorrow, budget for credits from Wednesday, and read the restoration promise as genuine but undated.

The precedent will outlast the episode

Oppenheimer's hearing did not slow the bomb; it clarified who owned it. The AEC upheld his revocation on 29 June 1954, and Commerce lifted these controls on 30 June 2026, seventy-two years almost to the day. The Fable 5 affair will be remembered the same way: not for the jailbreak, but for the precedent that a US administration can and will switch off a commercial AI model as leverage.

Amodei spent years arguing the state should hold that power over dangerous machines. He was right that it would matter. What the past month taught him, and everyone downstream of American AI, is that the state decides what counts as dangerous, and the deciding can take one evening. The window on Fable 5 closes on Tuesday, at the end of its seventh day.