AI found the crypto flaw two years of experts missed — then killed the standard

Anthropic's Claude Mythos Preview improved the best-known attack on HAWK, a third-round NIST post-quantum signature candidate, after about 60 hours of semi-autonomous work. It found a hidden mathematical symmetry in HAWK's lattice — a nontrivial automorphism that experts had theorized could exist but never confirmed — and effectively cut the scheme's key strength in half.

The sting is who missed it: HAWK had survived two rounds of expert human review over two years, and the Anthropic researcher steering Mythos was not a cryptography expert. Two of the model's own worker agents argued about the idea — one rejected it as infeasible before the other proved it worked. The whole run cost about $100,000 in API calls.

Nothing deployed is at risk: HAWK was never a production standard, and Anthropic says that is the point — the result puts AI in the standards-review loop, stress-testing candidate algorithms before they secure the internet. HAWK's authors were given advance copies and coordinated disclosure through NIST.