Anthropic Says Claude Found New Attacks on HAWK and AES Variants

Anthropic says its Claude Mythos Preview model found a new key recovery method against HAWK, a candidate in NIST's post quantum signature review, in a research paper. The company also described a faster attack on a seven round research version of AES, while stating that neither result affects current production systems.

For HAWK, the method uses a mathematical symmetry in the scheme's lattice structure. The paper says the attack lowers estimated key recovery costs for HAWK-512 from 2^150 to at most 2^108 gates, and for HAWK-1024 from 2^288 to at most 2^182 gates. HAWK has not been standardized or deployed.

The AES work targets seven rounds, not the full AES-128 algorithm used in production. The model improved a meet in the middle attack and made it 200 to 800 times faster in the research setting. The attack still requires assumptions used for academic analysis, including about 2^105 chosen plaintext encryptions.

Researchers spent weeks validating the model's findings. The release also includes CryptanalysisBench, a benchmark for testing language models on cryptanalysis tasks.