Overview
- Anthropic disclosed Tuesday that its unreleased Claude Mythos Preview autonomously found a key‑recovery attack that halves the effective strength of the HAWK-256 challenge parameter by exploiting a previously unused lattice symmetry.
- The company also published work showing Claude produced a mathematical shortcut called the Möbius Bridge that makes the best known attack on a seven‑round test version of AES‑128 roughly 200 to 800 times faster, though that result targets a reduced research cipher and needs impractical chosen plaintexts.
- Each model discovery took about 60 model‑hours and roughly $100,000 in compute, while Anthropic researchers then spent several hundred hours on human verification before releasing papers, code and reproducibility artifacts.
- Anthropic said it followed responsible disclosure, notifying HAWK designers, coordinating with a NIST mailing list, and briefing government and industry partners, and it released CryptanalysisBench tools built with academic partners to let others test AI cryptanalysis.
- The findings do not affect deployed systems today but could force HAWK designers to raise key sizes and will press standards bodies and operators to speed verification, update parameters, and plan how to respond if future models find flaws with immediate real‑world impact.