Researchers at Anthropic utilized the Claude Mythos Preview model to autonomously discover mathematical flaws in cryptographic algorithms, moving beyond simple implementation errors to identifying fundamental algorithmic weaknesses. One significant finding involves an improved attack on HAWK, a post-quantum digital signature candidate. By identifying a previously unexploited nontrivial automorphism in the lattice used by HAWK, Mythos reduced the scheme's effective keysize by half—for example, reducing the expected cost of a full key recovery for HAWK-256 from 2^64 to 2^38. While this does not break the full cipher, it necessitates doubling key sizes to maintain intended security levels. Additionally, the model improved cryptanalysis on a 7-round version of the Advanced Encryption Standard (AES). By developing a "Möbius Bridge" fingerprinting algorithm that remains invariant to certain guesses, Mythos eliminated a factor of 256 in workload, resulting in attacks that are 200–800× faster than previous methods. These research breakthroughs, achieved with an estimated API cost of $100,000 each, demonstrate the potential for frontier AI to perform high-level cryptanalysis, though the specific results currently affect only candidate or reduced-round schemes rather than production systems.