Anthropic shuts down Fable 5 just days after its release.
At 5:21 on a Friday evening, Anthropic received a letter from the US government — and by the time it had complied, two of its models, Fable 5 and Mythos 5, were switched off. The order was an export-control directive, citing national securi
At 5:21 on a Friday evening, Anthropic received a letter from the US government — and by the time it had complied, two of its models, Fable 5 and Mythos 5, were switched off. The order was an export-control directive, citing national security authorities. It instructed Anthropic to suspend access to both models for any foreign national, whether inside or outside the United States, including the company's own foreign-national employees. The net effect, Anthropic said, was that it had to disable the models for every customer at once. Every other Anthropic model kept running; only these 2 went dark. What prompted the order is, by Anthropic's account, unclear. The letter included no specific details of the government's concern. The company's understanding is that the government believes it found a method of jailbreaking Fable 5 — getting it past its safety rules. Anthropic says it reviewed a demonstration of the technique and found something narrow: in essence, asking the model to read a particular codebase and fix its software flaws. That surfaced a handful of previously known, minor vulnerabilities — the kind, the company notes, that other publicly available models, including OpenAI's GPT-5. 5. That distinction sits at the heart of Anthropic's case. There are narrow jailbreaks, which work in one specific situation, and universal ones, which broadly defeat a model's safeguards. After thousands of hours of red-teaming — with the US government, the UK AISI, outside groups, and its own internal teams — Anthropic says no one has found a universal jailbreak. Its position is that perfect resistance probably isn't possible for any provider, so rather than promise an unbreakable lock, it built what it calls defense in depth: safeguards meant to keep jailbreaks narrow or expensive to produce, monitoring to catch attacks, and a thirty-day data-retention policy — one that carries real costs with customers — so it can study and shut them down. That approach, the company argues, keeps the model's risks comparable to those of systems already deployed across the industry. It adds that it has received no disclosure of a harmful result from any such jailbreak. Anthropic is complying with the order, but it disagrees with it. Recalling a model used by hundreds of millions of people over a single narrow jailbreak would, it warns, effectively halt new model deployments across the whole industry if it became the standard. The company is careful to say it is not arguing that governments should never intervene — it believes a government should be able to block an unsafe deployment, through a process that is transparent, fair, clear, and grounded in technical facts. Its complaint is simply that this action did not meet that bar. For now, Anthropic has apologized for the disruption, called the episode a misunderstanding, and said it is working to restore access — with more details promised within a day.