OpenAI 设计 AI Agent 抵御提示注入

How ChatGPT defends against prompt injection and social engineering by constraining risky actions and protecting sensitive data in agent workflows.