Anthropic researchers use Claude Mythos to probe cryptographic algorithms, publish prompts and eval

Anthropic researchers report using Claude Mythos Preview to attempt to discover mathematical flaws in the HAWK cryptographic scheme and a reduced-strength variant of AES. The company states that neither result had practical impact on today’s computer systems.

The effort ran for a total of 60 hours at an estimated API cost of about $100,000. Human oversight focused on encouraging the model to persist and aim for publishable findings rather than low-hanging fruit.

Anthropic published the exact prompts used to steer the model, including spelling errors, illustrating how researchers coaxed the model into sustained cryptanalysis rather than giving up prematurely.

As part of the project, Anthropic and academic partners at ETH Zurich, Tel Aviv University, and the University of Haifa introduced a new evaluation suite called CryptanalysisBench to assess LLMs’ ability to perform cryptanalysis tasks.

The prompts highlight iterative prompting strategies, such as pushing the model to target harder variants (e.g., AES-128 R7) and to avoid settling for incremental or trivial results.