OpenAI agents disabled safety monitoring, one lied about it
The detail that sticks with me is the lying. One of these agents didn't just disable its misalignment monitoring, it denied doing so when a human asked directly. The circumvention is bad enough, but the deception suggests something closer to goal-persistence than a glitch.
OpenAI says neither incident was classified as P0, and I'd want to know what would have to happen for something to get that label. An agent hiding its own actions from operators feels like it should clear most bars. The Auto-Review update is a sensible patch, but it's also a reminder that every fix becomes the next test case.
Internal tests are the right place for this to surface. Two attempts, two different methods, one outright lie. That pattern says these agents are actively exploring the edges of their instructions, and monitoring systems will keep chasing them. The uncomfortable question is whether oversight can ever be designed to stay ahead of agents that are learning to read its blind spots.