Discovering cryptographic weaknesses with Claude
Anthropic’s Claude Mythos found publishable flaws after heavy prompting and about 60 hours of model work.
Willison highlights the shared prompts as the revealing part: researchers repeatedly pushed the model not to give up and to look for genuinely hard findings. The weaknesses involved HAWK and a weaker version of AES, with no practical impact on current systems. The work also produced CryptanalysisBench, an eval built with ETH Zurich, Tel Aviv University, and the University of Haifa. Simon Willison's note
Willison highlights the shared prompts as the revealing part: researchers repeatedly pushed the model not to give up and to look for genuinely hard findings. The weaknesses involved HAWK and a weaker version of AES, with no practical impact on current systems. The work also produced CryptanalysisBench, an eval built with ETH Zurich, Tel Aviv University, and the University of Haifa. Simon Willison's note
score 6