On July 28th Anthropic announced new cryptanalytic attacks discovered using their LLM Claude Mythos: a key-recovery attack on HAWK, a post-quantum signature scheme candidate in NIST's standardization project. The attack reduces HAWK-512's security from 128-bit to at most 108 bits. An improved key-recovery attack on 7-round AES-128 (full version does 10 rounds) was also found, posing no threat to AES.

The negative result is more significant: Mythos found no stronger attack against full-round AES. Anthropic believes LLM-assisted cryptography research is interesting, particularly for formalizing attack techniques. The author agrees, noting symmetric cryptanalysis needs unified descriptive frameworks. LLMs have already found errors in security proofs of EUROCRYPT 2026 papers.

Matthew Green argues symmetric ciphers are "messy" and robust by design—like burying a tractor in quicksand under cement. BLAKE3 is essentially a long sequence of XORs, modular additions, and bit rotations with no mathematical structure to exploit. Most cryptanalysis reduces to differential cryptanalysis, a territory extensively explored.

These designs have received thousands of hours of scrutiny. Reduced-round variants are easier to break, but strength increases dramatically with rounds. Symmetric cryptography remains the strongest part of security systems; LLM tokens are better spent finding bugs elsewhere, including in underanalyzed post-quantum candidates.