Discovering cryptographic weaknesses with Claude

Anthropic researchers used an advanced AI model called Claude Mythos to discover mathematical flaws in the HAWK cryptographic scheme and a reduced version of AES, though these findings do not practically impact current computer systems.

Achieving these results required specific and persistent prompting because the AI initially assumed the tasks were impossible to solve. Human intervention was necessary to push the model past roadblocks, discourage it from giving up, and steer it toward publication-worthy research rather than easy targets.

The entire process involved sixty hours of work from the Mythos Preview model, resulting in an estimated API cost of one hundred thousand dollars. This research was conducted in partnership with academic institutions including ETH Zurich, Tel Aviv University, and the University of Haifa, leading to the creation of a new evaluation framework called CryptanalysisBench to test whether large language models can perform cryptanalysis.