Anthropic's Claude Mythos Preview finds cryptographic weaknesses in HAWK and reduced-round AES

Discovering Cryptographic Weaknesses with Claude

Anthropic's red team used Claude Mythos Preview to improve the best-known attack on HAWK, a NIST third-round post-quantum signature candidate, cutting its effective key strength in half after just 60 hours of work. They also found a new attack on a reduced-round version of AES that is 200–800× faster than previous methods. Neither result affects production systems: HAWK isn't deployed, and the AES attack doesn't break the full cipher. Each finding cost roughly $100K in API fees and was achieved mostly autonomously. Anthropic followed responsible disclosure, notified HAWK authors and NIST, and partnered with ETH Zurich, Tel Aviv University, and University of Haifa to release CryptanalysisBench.

Why it matters: Anthropic used its own model to break crypto algorithms — halving HAWK's strength and speeding up AES reduced-round attacks by hundreds of times. Both numbers are solid. But pure crypto research is distant from most AI practitioners' daily work, so R axis missed, keeping the s...