Anthropic Says Claude Mythos Found HAWK, 7-Round AES Flaws

A single Anthropic researcher, working with an unreleased model called Claude Mythos Preview for roughly 60 hours, reportedly found a mathematical shortcut against HAWK, one of the post-quantum digital signature candidates still in NIST's standardization process. The find is not an implementation bug. It is a previously unexploited symmetry in HAWK's lattice structure, and it drops the expected attack cost on HAWK-256 from roughly 2^64 operations to 2^38. That is close to a practical break for the small-parameter version. HAWK is not deployed anywhere, which is the honest reason nobody has to patch anything today.

The second finding is more of a stress test than a scare. Given about 3 days of autonomous work, on the order of a billion output tokens, and another six figures in API spend, the same model produced a technique it calls the Möbius Bridge, which speeds up the best known attack on 7-round AES-128 by 200 to 800 times depending on how you measure it. Full AES has 10 rounds, and the 10-round cipher remains unbroken. Round-reduced attacks are how cryptographers probe cipher margins, so this belongs on the research pile, not in a CVE.

What makes the write-up interesting to non-cryptographers is the economics. Two publishable cryptanalysis results, each for about $100,000 in inference, on schemes that had already been picked over by human experts. Anthropic notes that a lot of the wall-clock time was researchers validating the model's output, not the model doing the math, and that Mythos needed a 'good amount of prompting' before it stopped concluding the problems were impossible. Preliminary results are mentioned on LEA, Serpent-128, Salsa20, Poseidon, and SHA-1, some more consequential than others.

The honest caveat is that a single-vendor post about a single-vendor unreleased model is not the same as independent replication. Anthropic says it consulted academics, shared advance copies with US government and industry partners, and coordinated public disclosure with HAWK's authors on the NIST mailing list in June, but the write-up does not name the outside cryptographers who checked the work or say whether HAWK's team accepts the analysis. Take the specifics as reported, not settled.

If this line of work generalizes, the winners are standards bodies and internal security teams who now have a plausible way to run AI-assisted cryptanalysis over borderline candidates before they get baked into products. The bottleneck moves from finding attacks to validating them, which is a much better problem to have than the one we had last week.