US order forces Anthropic to pull Fable 5 and Mythos 5 for all users worldwide

Fable 5, jailbreak fears, and the security debate

The underlying concern within the US government, according to reporting, centered on a possible jailbreak that could bypass the model's built-in safety mechanisms. Fable 5 is designed to refuse prompts in biology, chemistry, cybersecurity, and knowledge distillation — the practice of training one AI model using the outputs of another. Anthropic has said that no known method exists to make Fable 5 do things that could not also be done with rival models such as ChatGPT 5.5, and that jailbreaks have been tested successfully at Heise on other AI systems as well.

The Mythos base model had already drawn attention for its ability to identify dangerous software vulnerabilities, including flaws that had gone undetected for decades. That capability has been used by US authorities and selected companies to find security gaps, but from the outset there were concerns the same technology could function as a cyber weapon. Anthropic said it initially withheld Mythos from a larger audience precisely because the model appeared too capable of exploiting software vulnerabilities.

Silent throttling and a thin market for alternatives

Complicating the picture, Anthropic also admitted that it had secretly and massively throttled Fable 5 usage by automatically switching users in the background to the weaker Claude Opus 4.8. The automatic switching, which Heise's video magazine c't 3003 reported happened quite frequently, was not visible to subscribers. Anthropic justified the practice as a protective measure against knowledge distillation — saying Fable 5 is designed to detect when it is being used to train another model and switch itself off in response. The company recently announced it would become more transparent about the issue, and had separately drawn sanctions from the US military for prohibiting use of its models in autonomous weapons systems.

The shutdown leaves customers with limited near-term alternatives. Model families such as Mistral currently cannot keep up with products from Anthropic or OpenAI, according to the article, and Fable 5 is reported to be significantly stronger than Opus 4.8 or ChatGPT 5.5 in Anthropic's own benchmark comparisons. In its own words, Anthropic said the move amounts to the best advertisement possible for alternative AI providers — a dynamic that may now be tested in the marketplace as users look for substitutes.

Anthropic said it would restore access to the models as soon as possible. For now, both Fable 5 and Mythos 5 remain offline for all customers, regardless of nationality or location, while the company works with US authorities to resolve what it characterized as a regulatory misunderstanding.