External · Stanford UIT · intermediate

Responsible Agentic AI

Stanford's guidance for AI agents that act on your behalf: they inherit your permissions, can act at scale and can misread a goal, so use least privilege, read-only access unless writes are essential, separate service accounts and human approval for consequential actions. Every agent must be classified into a low, moderate or high risk tier and a deployment category (personal, internal workflow or external-facing) before it is deployed.