Simon Willison's lethal trifecta warns that an agent with access to private data, exposure to untrusted content, and a way to communicate externally can be made to leak that data. Meta's Agents Rule of Two turns the idea into a design rule: an agent session should hold at most two of three risky properties unless a human approves.
Read in this order
- ArticleThe lethal trifecta for AI agents: private data, untrusted content, and external communicationStart here: the original framing, short enough to share with a team.
- ArticleMeta: Agents Rule of Two: A Practical Approach to AI Agent SecurityThe Rule of Two as a practical design rule for agent sessions.
- PaperDesign Patterns for Securing LLM Agents against Prompt InjectionsConcrete design patterns that restrict agents so injection cannot cause harm.