Discovering cryptographic weaknesses with Claude

Anthropic used Claude to discover mathematical weaknesses in cryptographic algorithms—testing it on HAWK encryption and a simplified version of AES. Simon Willison's Weblog reports the research took sixty hours of computational time at an estimated cost of one hundred thousand dollars, and required extensive human guidance to keep the model focused on meaningful discoveries rather than trivial findings. The article shares the prompts used, showing how much encouragement Claude needed to persist through the analysis. The crucial point: neither vulnerability threatens real-world systems in use today, but the work demonstrates how large language models can contribute to specialized cryptographic research.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton