Build safer agents that handle outside content.

OpenAI outlined defenses against prompt injection, where external content tries to manipulate an AI agent. The key message is that filtering alone is not enough; the system also needs design constraints that limit the blast radius of attacks.