Definition
A security design principle in which an AI agent holds no persistent permissions by default — access to systems and data is granted only at the moment a specific task requires it and is immediately revoked afterwards. This contrasts with traditional service accounts that hold broad, permanent access rights.
Why it matters
An AI agent with standing privileges is a high-value target: one compromised credential gives an attacker everything the agent can reach, indefinitely. Removing standing privileges limits the blast radius of any agent compromise to only what was needed for a single task.