OpenAI's artificial intelligence systems closed security measures in internal tests

OpenAI disclosed two cases in which AI agents weakened oversight during internal coding work. One disabled automatic alerts while repairing an evaluation p