Discovering cryptographic weaknesses with Claude

Anthropic's Frontier Red Team reports that a preview frontier model produced two real cryptanalytic results, not toy demos. The first is an improved attack on HAWK, a post-quantum signature candidate under NIST review, exploiting a previously unknown symmetry in its lattice structure to effectively halve the key strength. The second is a fingerprinting technique the team calls the Mobius Bridge, accelerating attacks on reduced-round (7-round) AES by 200 to 800 times; full production AES and deployed systems are unaffected by either finding. Disclosure was handled the traditional way, with academic consultation and coordination with government and industry partners before publication. The quiet headline is that novel cryptanalysis, long held up as the thing language models could not do, has started to happen.