Claude finds mathematical flaws in HAWK and reduced AES

Anthropic published Discovering cryptographic weaknesses with Claude on July 28: researchers used Claude Mythos to find mathematical flaws in the HAWK signature scheme and in a weakened version of AES. Anthropic's own caveat, quoted up front: "neither of these results has a practical impact on today's computer systems."

The artifacts matter more than the headline claim. Anthropic released a cryptography-research-demo repo with the setup, and — as Simon Willison notes — the actual prompts used, spelling mistakes included. His pick of the litter is the instruction working around the models' learned defeatism: "the models tend to think it is impossible."

#Why it matters

Cryptanalysis is an unusually clean testbed for frontier-model research claims: results are verifiable, and the weakened-primitive framing lets you calibrate capability without publishing a live break. It also lands in the middle of the ongoing argument over whether models can do novel mathematics or only recombine known attacks — and the published prompts are the part that lets outsiders check. The HN thread (226 points, 177 comments) is doing exactly that.

Sources: Anthropic, Simon Willison, Repo, HN discussion