Anthropic's Claude Mythos Breaks Post-Quantum Crypto That Stumped Human Experts
AI model discovers mathematical flaws in NIST candidate HAWK and improves AES cryptanalysis in 60-hour autonomous research run.
Anthropic researchers using a preview of their Claude Mythos model have discovered mathematical vulnerabilities in two major cryptographic algorithms, marking a shift from AI finding implementation bugs to identifying fundamental weaknesses in the algorithms themselves.
The AI identified a new attack on HAWK, a post-quantum digital signature scheme under consideration by the National Institute of Standards and Technology (NIST). The attack exploited a previously unused symmetry—a nontrivial automorphism—in HAWK's lattice structure, reducing the expected cost of breaking HAWK-256 from 2^64 to 2^38 operations and effectively halving the scheme's security margin.
HAWK had survived two years of human expert review as a NIST third-round candidate. The agency is racing to replace current standards like RSA and ECDSA, which are vulnerable to future quantum computers.
In a second finding, Claude Mythos invented a fingerprinting technique called the "Möbius Bridge" that improves the best-known attack on 7-round AES-128 by a factor of 200 to 800 times. The AES finding targets a reduced-round version and does not threaten the full 10-round AES used in production systems.
The HAWK discovery required approximately 60 hours of semi-autonomous work with an estimated API cost of $100,000. While these specific attacks do not currently threaten deployed systems, the findings suggest that frontier AI models could drastically accelerate the discovery of vulnerabilities in the foundations of digital security.
The development raises questions about the future of cryptographic review processes that have traditionally operated at human speed. Anthropic has begun broader audits of cryptographic schemes and plans to host academic workshops on the role of language models in security research.
Security researchers note that while HAWK remains a candidate rather than a deployed standard, the episode demonstrates that AI-assisted cryptanalysis has moved beyond theoretical possibility into practical territory. The cost barrier—$100,000 for a single discovery run—remains significant but may decrease as models become more efficient.
NIST has not yet commented on whether the HAWK findings will affect its standardization timeline. The agency continues to evaluate post-quantum candidates as the cryptographic community prepares for the eventual arrival of fault-tolerant quantum computers.