Anthropic has publicly disagreed with the US government directive that forced it to suspend access to Claude Fable 5 and Mythos 5, saying in a follow-up statement that the decision rests on a misunderstanding of a limited jailbreak rather than on any new or unusual capability in the models.
The company says it is complying with the order while it works to restore both models as quickly as possible.
What Anthropic says about the jailbreak
According to Anthropic, the technique cited by US officials as the basis for the action is not a universal jailbreak and does not unlock any extraordinary capability. The company says the method surfaced only a handful of minor, previously known software vulnerabilities, and that the same abilities are readily available in other publicly accessible AI models.
Anthropic also says that Fable 5 went through thousands of hours of security testing before release, involving US officials, the UK AI Security Institute, private organizations and the company’s own internal teams. Despite all of that testing, the company says, no comprehensive method was found that could broadly disable the model’s safety limits.
Capabilities “already in other models”
The statement goes further, suggesting that the capabilities described in the government’s report already exist in other frontier systems, including GPT-5.5, and that cybersecurity professionals use those same capabilities every day for defensive work.
Anthropic warned that if models can be pulled from the market on the strength of a narrow, non-universal jailbreak, nearly every advanced AI deployment could be put at risk in the future, slowing the whole industry.
Compliance, with objections
The company repeated its apology to customers affected by the suspension and said it is following the government’s instruction. At the same time, it described the decision as the product of a misunderstanding rather than of the technical facts, and said it is continuing efforts to bring Fable 5 and Mythos 5 back into service.
The suspension, which took effect in mid-June, was widely reported after Anthropic disabled access to its two most capable models in response to the US directive. As of the statement, the government had not publicly detailed its reasoning beyond the reported jailbreak findings.
