A security approach that grants nothing automatically: every person, device and program must prove who it is and that it is allowed, each time it asks for something. Older security treated anything inside the company network as safe. Zero trust assumes attackers may already be inside, so it checks every request and gives the least access needed. For AI systems, the same thinking applies to agents and the tools they call.
Technical view
Popularised by Forrester analyst John Kindervag in 2010 and codified by the United States National Institute of Standards and Technology (‘NIST’) in Special Publication 800-207 (August 2020), which moves defences from network perimeters to users, devices and individual resources. Authentication and authorisation happen per session against policy, using identity, device state and context. Applied to AI, it means treating models, agents and tool servers as distinct identities with scoped, short-lived credentials, never treating model output as an authorisation decision, and handling retrieved content and tool descriptions as untrusted input. The costs are policy complexity and friction for users.
Related
Part of AI Terminology Index, section: Security and Privacy.