Anthropic's Claude Mythos finds cryptographic weaknesses in HAWK and AES
Anthropic researchers used Claude Mythos to discover mathematical flaws in the HAWK cryptographic algorithm and a reduced-round version of AES. The findings, detailed in a blog post by Simon Willison, have no practical impact on current computer systems. Mythos Preview worked for 60 hours at an estimated API cost of $100,000. The main human interventions involved prompting the model to persist and aim for publishable results, with prompts including misspellings like "find something that worth publishing." The research repo is available online.
Key facts
- Anthropic researchers used Claude Mythos to find flaws in HAWK and a weaker AES version.
- Neither finding has practical impact on today's computer systems.
- Mythos Preview worked for 60 hours with an estimated API cost of $100,000.
- Human interventions encouraged the model not to give up and to find publishable results.
- Prompts included misspellings such as 'find something that worth publishing'.
- The article was posted by Simon Willison on 28th July 2026.
- The research repo is available online.
- Claude Mythos is an AI model developed by Anthropic.
Entities
Institutions
- Anthropic
- Simon Willison