AI Signal 194
Discovering cryptographic weaknesses with Claude
LLMs can now contribute to cryptanalysis research, but only with heavy human prodding and at significant cost—an estimated $100,000 in API spend over 60 hours. The work also produced CryptanalysisBench, a new eval for measuring LLM cryptanalysis ability, developed with ETH Zurich, Tel Aviv University, and University of Haifa.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Claude Mythos identified mathematical flaws in HAWK and a weaker AES variant, though neither has practical impact on today's systems.
The model tended to assume problems were impossible and stop trying, requiring repeated human intervention to persist toward publishable findings.
The project produced CryptanalysisBench, a new evaluation benchmark for LLM cryptanalysis capabilities, developed in partnership with three universities.
THE CLUSTER