Anthropic's Claude cracks new attacks on post-quantum HAWK and reduced AES

Anthropic’s Frontier Red Team used a research build called Claude Mythos Preview to find genuine mathematical flaws in cryptographic algorithms, not just implementation bugs in crypto libraries. The headline result targets HAWK, a lattice-based digital signature scheme still in the third round of NIST’s post-quantum standardization. In roughly 60 hours the model improved on the best known attack by exploiting a previously unused symmetry (a nontrivial automorphism) in HAWK’s lattice, effectively halving the scheme’s key strength. That matters because restoring security would require doubling key sizes, which erases much of what made HAWK an attractive candidate in the first place. A second result sped up the best known attacks on a round-reduced variant of AES by 200-800x by eliminating one of the guesses an attacker must make.

Neither finding threatens anything in production: HAWK isn’t deployed, and the AES work only touches a deliberately weakened version, not the full cipher. The significance is the demonstration that a frontier model can do real cryptanalytic research — literature review, mathematical reasoning, computational experiments in Python and Sage, and end-to-end verification of its own results. The HAWK attack was developed semi-autonomously with light human direction from a researcher who wasn’t a cryptographer, while the AES attack was found fully autonomously inside a purpose-built scaffold. Each result ran to about $100,000 in API costs.

Anthropic frames this as cryptanalysis working as intended — stress-testing algorithms to build trust — and followed responsible disclosure, sharing the HAWK attack with its authors and the NIST mailing list and briefing government and industry partners in advance. To let others probe LLM cryptanalysis, it partnered with ETH Zurich, Tel Aviv University, and the University of Haifa to release CryptanalysisBench, a benchmark bundling multiple ciphers for evaluation. The broader implication: AI is becoming a practical tool for finding weaknesses in cryptographic standards both before and after deployment.