How to Approach Security in the Era of AI Agents

### Executive Summary: This guidance outlines five critical security controls for AI agents—strict separation of authentication/authorization from model outputs, treating model outputs as untrusted with rigorous validation, executing generated code in isolated sandboxes, comprehensive observability and audit logging of agent decisions and tool invocations, and continuous security testing and red teaming—to reduce risks from unintended actions, overprivileged tools, hallucinations, and adversarial misuse while maintaining human oversight.