Anthropic's AI Model Finds Critical Flaw in NIST Post-Quantum Cryptography Candidate, Prompting Withdrawal

An advanced AI model developed by Anthropic discovered a previously unknown mathematical vulnerability in HAWK, a leading candidate for new post-quantum digital signature standards from the U.S. National Institute of Standards and Technology (NIST), forcing the cryptography scheme's withdrawal from consideration within 24 hours of the disclosure. Claude Mythos, an experimental large language model, identified a nontrivial automorphism—an exploitable symmetry in HAWK's underlying lattice structure—after approximately 60 hours of directed research. The finding reduced HAWK-256's practical security strength by roughly half: from an intended 128-bit security level to approximately 108 bits maximum, while recovering a secret key on commodity hardware in under four hours.