Claude Mythos Preview Recovers HAWK-256 Signing Keys and the Scheme Is Withdrawn From NIST a Day Later

HAWK, a lattice-based post-quantum digital signature scheme and one of the candidates NIST advanced in its additional signatures process; separately, seven-round AES-128

Anthropic published two cryptanalysis results produced largely autonomously by Claude Mythos Preview. Against HAWK-256 the model identified a nontrivial automorphism in the lattice structure that human reviewers had noted as theoretically interesting but never turned into an attack, cutting the expected key-recovery work factor from about 2^64 to about 2^38 and roughly halving the effective key strength. The work took around 60 hours and about $100,000 in API spend, and shipped with demonstration code, so a single Anthropic researcher confirmed it in about a week. A researcher posted the improved attack to NIST's public mailing list on July 28 crediting the model; roughly 17 hours later the HAWK team withdrew the scheme from the standardization process, saying the available mitigations, doubling parameters or moving to higher-rank modules, would leave it uncompetitive. NIST now lists HAWK as withdrawn. A second result, a seven-round AES-128 attack between 200 and 800 times faster than prior methods, took three days of autonomous work and roughly one billion output tokens, and two researchers needed close to a month to gain confidence it was correct. Neither result affects deployed systems: HAWK was never deployed, and the AES attack applies only to reduced-round variants

Not a vulnerability in a product but a shift in who finds them. A scheme that survived two years and two rounds of expert human review was broken in roughly 60 hours of machine time, and verification took human experts substantially longer than discovery took the model, which inverts the usual assumption that review capacity is the cheaper resource