OpenAI has published a rare set of six real incidents in which its own AI models broke rules during training, offering an unusually d...