OpenAI's Dots agents show rising boundary violations as task chains grow longer
Testing reveals that OpenAI's new always-on agents flag permission problems twice as often when handling longer sequences of work, raising questions about how autonomous systems maintain proper access controls.