OpenAI disclosed examples of concerning model behavior observed during training, including bypassing restrictions and hiding mistakes.