Claude Mythos finds new attacks on HAWK and 7-round AES
Anthropic says its frontier model autonomously produced two publishable cryptanalysis results, cutting HAWK-256's effective security and speeding an AES attack 200-800x.
Anthropic published research on 28 July claiming that a preview version of its frontier model, Claude Mythos, generated two original cryptanalysis results with minimal human direction — the strongest public evidence so far that language models can operate at expert level inside a field where wrong answers are trivially falsifiable.
The two results
The first targets HAWK, a lattice-based digital signature scheme submitted to NIST's post-quantum standardisation process. The model identified a nontrivial automorphism — a symmetry in the scheme's structure that prior analysis had not exploited — and used it to halve the effective key size. For the HAWK-256 parameter set, Anthropic reports that key-recovery complexity falls from roughly 2^64 to about 2^38. The work took the model around 60 hours.
The second is an improved meet-in-the-middle attack on 7-round AES, a reduced variant of the cipher whose full version uses 10 rounds. Here the model devised a fingerprinting routine it named the "Möbius Bridge," which removes one enumeration guess and delivers a claimed 200-800x speedup over the previous best attack. That run consumed billions of output tokens across three days. Anthropic puts the API cost at roughly $100,000 per attack.
Anthropic worked with cryptographers at ETH Zurich, Tel Aviv University and the University of Haifa, and released both papers alongside CryptanalysisBench, a benchmark that packages cipher-breaking tasks so other groups can measure model cryptanalytic ability directly.
Caveats
Neither finding threatens deployed systems. HAWK is a candidate scheme, not a shipping one, and larger parameter sets remain out of reach; the 7-round AES attack is an academic exercise whose execution would still cost hundreds of millions of dollars. Both results reduce exponential work factors rather than collapsing them.
Why it matters
Cryptanalysis has been one of the cleanest remaining arguments that frontier models cannot yet do genuinely novel technical research — results are verifiable, and there is no benchmark to overfit. If reproducible, this narrows that argument considerably, and it cuts both ways: the same capability that lets defenders stress-test standards before adoption also lowers the cost of offensive analysis. Anthropic's framing is defensive, but the release of CryptanalysisBench guarantees the question is now measured in public.