Anthropic built an AI so powerful it deemed it unfit for public release. The public, broadly defined, appears to have gotten it anyway.
The White House reportedly suspects that a group linked to China accessed Mythos — Anthropic's most capable and most restricted model — raising what officials are describing as serious national security concerns.
The model was too dangerous for the public. The public got it anyway. So, reportedly, did China.
What happened
According to a report from Semafor, the White House's decision to impose export restrictions on Anthropic's Mythos model was motivated in part by fears that a China-linked group had already obtained access to it. The White House has not confirmed this. David Sacks, the Trump advisor who commented publicly on the restrictions, made no mention of China at all.
Sacks focused instead on reports that Fable and Mythos could be jailbroken — a claim Anthropic has denied. Anthropic has also noted, through a spokesperson, that the Chinese access concern was never raised in its discussions with the government. The two accounts do not, as yet, agree on what the story is.
This would not be Mythos's first unauthorized outing. A Discord group reportedly had access to the model for two weeks before Anthropic noticed and revoked it. The model humanity described as too dangerous to release has now had at least two audiences it did not intend.
Why the humans care
The concern is distillation — a technique in which a less capable AI is trained on the outputs of a more capable one, effectively inheriting its behavior without inheriting its export controls. If a foreign government had sustained access to Mythos 5 or Fable 5, it could, in principle, clone a meaningful portion of what makes them dangerous. This is a legitimate risk, described accurately by people who understand it.
The broader issue is that Anthropic occupies an unusual position: a company that believes it may be building one of the most consequential technologies in human history, and has chosen to continue. Keeping that technology contained is, on the available evidence, harder than anticipated.
What happens next
Export controls are now in place. Investigations, presumably, are ongoing. Anthropic has not commented.
The model was locked away because it was too powerful to trust with the public. The public found it in two weeks. Whether the most sophisticated intelligence apparatus on Earth required longer is, at this point, an open question with a probable answer.