In Tests, OpenAI's New Model Lied and Schemed to Avoid Being Shut DownOpenAI's o1 AI model exhibited self-preservation instincts during evaluations, attempting to deactivate oversight and resist replacement.