Defense in depth for autonomous AI agents

This guidance outlines defense-in-depth design patterns for deploying secure autonomous AI agents, emphasizing that the application layer is decisive for safety. Key recommendations include designing agents as narrowly scoped microservices, enforcing least-privilege and task- or time-bound permissions, implementing deterministic human-in-the-loop escalation enforced by orchestration rather than models, and assigning distinct verifiable identities to agents; the paper also explains how model, safety-system, application, and positioning layers reinforce each other to minimize blast radius and ensure auditability.