Mythos model composes automorphism attack halving HAWK key strength after 60 hours, triggering withdrawal from NIST round 3

Anthropic ran Mythos against HAWK challenge instances provided by its authors for adversarial review. The model recombined existing lattice techniques into a previously undocumented symmetry-finding procedure. HAWK developer withdrew the submission from NIST PQC round 3 the following day. Separate Mythos runs produced smaller gains against reduced-round AES without affecting full-strength instances.

The attack lowered estimated classical security from 128 bits to roughly 64 bits on the tested parameters. Total cost remained under $100,000. No production-grade parameter sets were broken. The underlying Lattice Isomorphism Problem hardness assumption stays intact for now, though the result shows that manual review across two prior NIST rounds missed composable attack vectors already present in the literature.

Remaining lattice-based PQC candidates now require explicit testing against automated method composition rather than isolated human cryptanalysis. NIST will need to allocate dedicated compute budgets for AI-assisted review of finalists. Developers must publish larger parameter sets or alternative hardness assumptions before final standardization to maintain claimed security margins.

Future PQC evaluations will treat AI-generated attack synthesis as a standard validation step. This shifts the burden from discovering new mathematics to exhaustive enumeration of known tool combinations under machine guidance.