Anthropic says its Claude Mythos Preview model has pulled off two notable cryptanalytic feats: a full key-recovery attack on HAWK-256, a test candidate for post-quantum signatures, and a 200- to 800-fold speedup for an attack on seven-round AES-128. Neither result threatens real-world systems, the company stresses, but both show how advanced AI can uncover mathematical shortcuts that humans had missed.

The HAWK attack exploits a symmetry hidden in the lattice underlying the signature scheme. Previous work had identified a path to recovery, but it required an automorphism that the AI discovered on its own. Anthropic’s implementation runs in under four hours on a 96-core server and reconstructs a signing key functionally equivalent to the original. The work factor for HAWK-256 drops from 2^64 to 2^38; for the larger, NIST‑standardization parameters HAWK-512 and HAWK‑1024, the estimated gate count falls but remains far beyond practical reach. Anthropic says the attack remains exponential and does not extend to other lattice‑based candidates.

The AES result targets a version of the cipher reduced from ten rounds to seven. Previous meet‑in‑the‑middle attacks required guessing 256 possible values at one stage; the AI generated an invariant fingerprint—called the Möbius Bridge—that eliminates that guess. The speedup is large but still requires an impractical number of chosen plaintexts, so no production AES‑128 software needs changes. Anthropic’s papers and verification code are public, and the company notes that the model initially refused to engage on the AES problem before human researchers pushed it to keep looking. The total API cost for both projects was about $100,000, but verification took researchers months.

Source: The Hackers News