Claude found new weaknesses in two cryptographic schemes

Anthropic says its Claude Mythos Preview model, working largely on its own, produced two genuine cryptanalytic results that had escaped years of expert review. The first targets HAWK, a post-quantum signature scheme that NIST is considering for standardization. Claude found an improved attack that cuts the effective key size roughly in half, dropping the expected cost of breaking HAWK-256 from about 2^64 operations to 2^38. Getting there took around 60 hours of model work and roughly $100,000 in API spend, which is a useful data point on what serious automated research actually costs today.

The second result is a new fingerprinting method, which Anthropic calls the "Möbius Bridge" algorithm, that speeds up attacks on reduced-round AES by 200 to 800 times over prior techniques. This only reaches 7 of the 10 rounds in AES-128, so full AES is not broken and nothing you use today is at risk. HAWK is not deployed either, and the AES work targets weakened variants that researchers study on purpose. Anthropic went through responsible disclosure and consulted government and industry partners before publishing.

The striking part is that a model surfaced real mathematical flaws in algorithms that many specialists had already picked over. The write-up is careful about scope, which makes it more convincing than the usual claims of AI doing science.

Why it matters

If you work in cryptography or security, this is a signal to fold AI assisted cryptanalysis into how you evaluate new schemes, because a model just found a result on a NIST candidate that human review missed. For everyone else the reassuring part is concrete: deployed encryption is untouched, and the findings went through disclosure first.