Two weeks. That is how long an unauthorised Discord group reportedly had access to Mythos, Anthropic's most powerful and closely guarded AI model, before the company noticed the breach and cut off the connection. That detail, uncomfortable enough on its own, has now been overshadowed by something potentially far larger: according to reporting by Semafor, the White House moved to impose export restrictions on Mythos partly because of suspicions that a group linked to the Chinese government had already gained access to the system. The implications, if confirmed, would be significant.

Mythos — along with a related model called Fable 5 — has been described by Anthropic itself as too dangerous for public release. Were a state actor to obtain sustained access, officials fear the consequences could reach well beyond commercial competition. Of particular concern is a technique known as distillation, in which a less capable "student" AI is trained on the outputs of a more advanced model in order to replicate its behaviour. Through distillation, even limited access to a frontier system could allow a rival to accelerate its own AI development considerably.

The White House has not publicly confirmed the China angle.