Claude Mythos Worked ~60 Hours Autonomously to Find Mathematical Flaws in HAWK and Weakened AES
Simon Willison·medium signal
Anthropic researchers used Claude Mythos to discover cryptographic weaknesses in the HAWK signature scheme and in deliberately weakened AES variants, with the model running roughly 60 hours of autonomous work. Human intervention was minimal and mostly consisted of encouragement not to give up and to push for publishable-quality findings rather than technical steering. As a data point on long-horizon autonomy this is more interesting than most benchmark numbers: the bottleneck reported was persistence, not capability.