TL;DR
Anthropic researchers report that Claude Mythos Preview found mathematical flaws in cryptographic algorithms themselves, not merely in how they are implemented. The model improved the best-known attack on HAWK, a candidate post-quantum digital signature scheme, in 60 hours — effectively halving its key strength after the scheme had survived two years and two rounds of expert human review. Neither result affects production systems.
The HAWK attack works by finding a previously unexploited symmetry in the lattice the scheme relies on. Earlier work had proved such a symmetry would permit an attack but had not established whether one existed in HAWK’s lattice. The practical consequence is that HAWK keys would need to double in size for the same security — which removes much of what made it attractive as a candidate in the first place. The attack is specific to HAWK and does not affect other NIST post-quantum candidates or lattice cryptography generally.
The second result concerns AES, adopted by NIST in 2001 and among the most scrutinised ciphers in existence. Working on a seven-round variant of the ten-round AES-128, the model eliminated one of the guesses an attacker must make, improving on previous best attacks by 200 to 800 times. The full cipher is not broken.
Method matters as much as outcome. One researcher worked alongside the model on HAWK; another built a scaffold that let it discover the AES attack autonomously. Each result cost roughly $100,000 (£75,000) in API charges. Anthropic followed responsible disclosure, sharing the HAWK finding with its authors in June and coordinating publication with the NIST mailing list, and has released CryptanalysisBench with academics at ETH Zurich, Tel Aviv University and Haifa.
Looking Forward
The reassuring framing — cryptography stress-testing working as intended — holds only while the capability stays expensive and disclosed. At $100,000 per result it is within reach of a well-funded adversary today, and the finding arrived late in a standardisation process specifically designed to catch flaws before deployment. For UK security teams and the AI Security Institute’s cyber-evaluation programme, the useful signal is the compression: two years of expert review, then 60 hours. Standards bodies now need review cycles that assume this tooling exists on both sides.