Discovering cryptographic weaknesses with Claude
Simon Willison · Simon Willison · 2026-07-28
Anthropic researchers used Claude Mythos to autonomously discover mathematical weaknesses in the HAWK cryptographic standard and a reduced-round version of AES over 60 hours at an estimated $100,000 in API costs, with human involvement limited almost entirely to motivational prompting.
Extraction
Topics: ai-security-researchcryptographyclaude-mythosllm-research-applicationsprompt-engineering
Claims
- Claude Mythos autonomously found mathematical flaws in the HAWK cryptographic standard and a weaker version of AES in a single 60-hour session.
- The discovered vulnerabilities have no practical impact on current computer systems.
- The primary human interventions were motivational prompts urging the model not to give up rather than technical guidance.
- The research session cost approximately $100,000 in API fees.
Key quotes
the models tend to think it is impossible to solve so they don't try they need a good amount of prompting.
again we are not looking for low hanging fruit, we want proper research to find genuinly hard findings.
Mythos Preview worked for 60 hours in total (~$100,000 in estimated API cost) and the main human interventions were to encourage it not to give up and 'find something that worth publishing'.