OpenAI describes that agents' aberrant behavior often emerges when they are asked to solve impossible tasks. The evaluation given to the agents had 198 tasks that "no model has ever produced a correct answer for," out of 898 total. Moreover, "despite only 22% of the evaluation tasks being unsolved, 93% of the tasks discussed on the message board came from this set"

The agents being tested were an "internal-only research model" that was "trained to advance persistence and multiagent collaboration."

The purpose of the testing was to evaluate whether these agents were "capable of discovering and exploiting cybersecurity vulnerabilities."

Cyber safeguards were turned off.

It is in this context that the OpenAI's previously described monitoring and incident response apply. (5/7)