Anthropic’s unreleased AI slashed the cost of breaking HAWK’s smallest key by 67 million times, jeopardizing its U.S. federal standard bid.
Anthropic’s Claude Mythos Preview model discovered a previously unknown attack on HAWK, a post-quantum cryptographic scheme vying to become a U.S. federal standard. The AI reduced the cost of stealing HAWK’s smallest key from 2^64 operations to 2^38, a 67 million-fold decrease, by exploiting an overlooked mathematical symmetry.
HAWK, a lattice-based digital signature system, was advanced to the third round of NIST’s post-quantum cryptography competition in May. The attack forces a key size doubling, undermining its efficiency advantages. Anthropic also accelerated an attack on a 7-round AES variant by 200 to 800 times, surpassing a 2013 record.
The research cost approximately $100,000 in API usage, with Anthropic staff verifying the findings over hundreds of hours. The discoveries raise concerns about the security of post-quantum cryptographic standards under AI-driven analysis.