Least privilege for AI agents

Least privilege — grant only what's needed, nothing more — is the single most effective control for autonomous agents. Warrants make it practical.

From principle to mechanism

Everyone agrees agents should have minimal access; the hard part is expressing and enforcing "minimal" for a program that decides its own actions. A warrant makes minimal concrete: exact resources, exact actions, a spend ceiling, an expiry, holder-binding. Anything not granted is denied, and unknown caveats fail closed.

Minimal, and it stays minimal

Least privilege usually erodes as systems grow and permissions accrete. Attenuation reverses that pressure: authority can only shrink as it flows to sub-agents, so a deep multi-agent system tends toward less privilege downstream, not more.