Simon Willison • 7/28/2026

Discovering cryptographic weaknesses with Claude

This article details how Anthropic researchers leveraged Claude Mythos, an advanced AI model, to discover mathematical flaws in the HAWK algorithm and a weakened version of AES, though without practical impact on current systems. It highlights the prompts used to encourage the model to persist in finding publishable research, costing approximately $100,000 in API fees over 60 hours. The work also introduces CryptanalysisBench, a new evaluation benchmark developed with ETH Zurich, Tel Aviv University, and University of Haifa, demonstrating AI's potential in cryptanalysis.

0 comments

#llm #AI Research #Ae