Anthropic says AI found attacks on HAWK-256 and reduced-round AES

Anthropic said its Claude Mythos Preview model uncovered new cryptographic attacks in findings published on July 28, 2026, underscoring how frontier AI systems are becoming useful tools for expert-level cryptanalysis. The company said the most significant result found a previously unused symmetry in HAWK, a post-quantum digital signature candidate in NIST’s standardization process, reducing the expected cost of full key recovery for HAWK-256 from 2^64 to 2^38 operations. Anthropic said the HAWK work took about 60 hours, with each of the two main studies costing roughly $100,000 in API credits. For HAWK-512 and HAWK-1024, it said the attack remains practically infeasible, though maintaining the same security level would require roughly doubling key sizes and eroding one of HAWK’s main advantages: compactness. Separately, Anthropic said Claude Mythos Preview found an attack on 13 rounds of the Korean LEA block cipher standard that could recover the key in under an hour on a modern desktop computer, while stressing that the full cipher uses 24 rounds and that current systems are not at risk. Anthropic said the findings were disclosed in advance to the algorithm authors, U.S. government agencies and industry partners, with the HAWK result coordinated through NIST’s mailing list. The broader significance for crypto is not that Bitcoin or deployed software was affected, but that AI-assisted cryptanalysis may be making it faster and cheaper to identify weaknesses in candidate algorithms before they are adopted. Anthropic and academic collaborators also introduced CryptanalysisBench, a 191-task benchmark spanning six families of cryptographic primitives, and said tested frontier models solved 65.3% to 85.7% of level-one tasks, while no model exceeded 9% on full schemes with no published practical attacks.