The only way to expose GPT-4o's system prompt, for example, is through a prompt injection attack. And even then, the system's output can't be trusted completely. However, Anthropic, in its ...