Anthropic’s Claude Mythos Preview model has autonomously discovered significant mathematical vulnerabilities in the HAWK post-quantum digital signature scheme and reduced-round AES, marking a milestone in AI-driven cryptographic research.

Key Points

• Claude Mythos Preview identified a mathematical flaw in HAWK, a NIST post-quantum candidate, effectively halving its key strength for the HAWK-256 variant.
• The AI achieved a 200 to 800 times faster attack on a reduced-round version of AES using a novel technique called a Möbius Bridge.
• Research was conducted semi-autonomously over several days, with the HAWK discovery requiring approximately $100,000 in API costs and 60 hours of processing.
• The model also developed a practical attack against 13-round LEA, recovering keys in under an hour using fewer than 2^30 encrypted plaintexts.
• Anthropic followed responsible disclosure protocols, notifying HAWK authors in June before coordinating a public release with the NIST mailing list.

Why it Matters

This development signals that frontier AI models can now perform high-level cryptographic research that surpasses the capabilities of human experts. As AI accelerates the discovery of complex security flaws, the industry faces a growing bottleneck where human researchers struggle to validate and remediate vulnerabilities at the speed of machine-generated output.