OpenAI wants to detect dangerous patterns across multiple AI interactions without letting its employees read customer prompts. Here is how Private Safety Processing and Zero Data Retention are supposed to work.