Definition
A security model where AI agents and automation systems hold no default or persistent permissions; instead, they request and receive time-limited, task-specific credentials only when needed and are revoked immediately after. This contrasts with legacy approaches where agents run under shared service accounts with always-on permissions.
Why it matters
AI agents operate at machine speed and can be compromised or misdirected before humans notice. ZSP prevents a compromised agent from becoming a persistent backdoor—it can only damage systems during its narrow window of delegated access, reducing the blast radius of agent compromise.