No model reliably separates instructions from data, so runtime defense is about limiting what injected text can cause. The practical rule is to avoid giving one session private data, untrusted content and an outbound channel at the same time, then layer detection and isolation on top. Assume some injections will land and size the blast radius accordingly.
Read in this order
- ArticleThe lethal trifecta for AI agents (Simon Willison)Start here: the three-part rule that frames every runtime defense.
- ArticleAgents Rule of Two: A Practical Approach to AI Agent Security (Meta AI)Meta's Rule of Two, a practical way to apply the trifecta to designs.
- Open sourceAI Security Labs: Runtime DefenseHands on: toggle gateway controls in a simulated email agent and watch exfiltration fail. Needs Docker and local models.