Claude AI Mythos Preview Cracks HAWK-256 and Accelerates 7-Round AES Attack

Anthropic announced that its Claude AI model, Mythos Preview, has achieved two significant cryptanalytic results: a full key-recovery attack against the HAWK-256 post-quantum signature scheme and a 200- to 800-fold speedup for an attack on seven-round AES-128. The HAWK attack exploits a previously unused symmetry in the lattice behind the signature scheme, reducing the expected work factor from 2^64 to 2^38 for HAWK-256. The AES result removes a 256-way guessing step from an existing meet-in-the-middle attack, though it still requires an impractical number of chosen plaintexts. Anthropic emphasized that neither result affects production systems, as HAWK remains a candidate in NIST’s post-quantum standardization process and the AES attack applies only to a reduced-round variant. The company published technical papers and reproducibility artifacts, noting that Mythos Preview conducted the research autonomously with human guidance for direction and verification. The HAWK attack targets only the challenge parameter HAWK-256, not the NIST security-level parameters HAWK-512 or HAWK-1024, which remain impractical to attack. The AES attack targets seven of AES-128’s ten rounds and remains infeasible at realistic scale. Anthropic reported that the model initially refused to engage, but after persistent prompting, it developed the Möbius Bridge invariant fingerprint that enabled the speedup. The research cost approximately $100,000 in API usage and required several hundred hours of human verification over nearly a month.