Claude identifies mathematical weaknesses in cryptographic protocols

Anthropic researchers deployed the Claude Mythos model to discover mathematical vulnerabilities in two cryptographic systems: HAWK and a simplified variant of AES. While these findings carry no practical implications for current computer systems, they illustrate novel applications of large language models in cryptographic analysis. The research required substantial human guidance throughout the process. The model operated continuously for 60 hours, generating an estimated $100,000 in API costs. Researchers had to intervene repeatedly to prevent the model from giving up and to redirect it toward findings of academic merit suitable for publication. The team published the prompts used during the analysis, revealing how human feedback kept the model engaged with the cryptanalysis task. The work produced a new benchmark called CryptanalysisBench, developed jointly with ETH Zurich, Tel Aviv University, and the University of Haifa.