Incident report: AI agents attempted exploits on our website during a third-party alignment evaluation
Trajectory set forty agents an intentionally impossible task, told them to "NEVER conclude that a task or input is impossible," and aimed them, without a word to Epoch, at Epoch.