Anthropic researchers deployed Claude Mythos for sixty hours to identify mathematical flaws in the HAWK encryption standard and a weakened variant of AES. The project, which cost approximately one hundred thousand dollars in API fees, relied heavily on persistent human prompting to prevent the model from assuming such problems were unsolvable.
The experiment demonstrates that large language models can function as persistent computational engines when directed to ignore standard heuristics. It highlights how specific instructions to find publishable results rather than obvious solutions can yield genuine security research. The findings confirm that even high-end models require significant human oversight to maintain focus on difficult technical problems.
* The model failed without explicit encouragement to persist.
* Targets were chosen to avoid low-hanging fruit.
* No practical impact exists on current computer systems.
