Agent Security
AI agent identity, authentication, and authorization
An AI agent is a software principal. It needs its own identity, bounded permissions, provenance, and a path back to human authorization.
Knowledge graph
Authoritative introductions to digital trust. No unverified product claims, no thin keyword pages.
All topicsThis page
Agent security
Last reviewed 2026-08-21
Next
Give agents identity, scope, evidence, and revocation — then require human approval for high-risk tools.
Request a consultationWho this is for
Teams building or governing autonomous agents that call APIs, move money, or change systems.
Problem
Agents often run as the user, with the user's token, and no distinct accountability.
What to do next
Give agents identity, scope, evidence, and revocation — then require human approval for high-risk tools.
Knowledge topics
LibraryAgent Security
What is an AI agent identity?
An agent identity is a unique principal for the agent instance or agent class, separate from the human who delegated to it and from the machine it runs on.
Agent Security
How should agents authenticate?
Agents should use workload or machine credentials to call APIs — not a copied user password. User context, when needed, should be a delegated grant, not a shared session cookie.
Agent Security
How should agents be authorized?
Authorization should be least privilege per tool: which APIs, which resources, which limits, which time window. High-impact tools should require human authorization, not only a system role.
Agent Security
Provenance and audit
Agent provenance records which agent acted, on whose behalf, with which policy version, and which human (if any) approved. Without that trail, incident response cannot assign accountability.
Agent Security
Revocation and liability
Organizations should be able to revoke an agent's grants immediately. Legal liability for autonomous actions is jurisdiction-specific and is not determined by this page.
Agent Security
Where this meets Keyra
Keyra publishes Human Authorized Intelligence™ and operates agent-related admin and marketplace surfaces. Public educational pages do not invent SDK methods. Use developer.keyra.ie for implementation contracts.
Questions
What people ask first.
Plain answers. No product claims that have not been published elsewhere on Keyra.ie.
How can an AI agent prove which human authorized it?
By presenting a grant that names the human, the scope, and the time — preferably as evidence that can be verified later, not only a log line.
How can AI actions be non-repudiable?
By binding the action payload to keys controlled under a defined custody model, plus retention of the authorization record. See Non-Repudiation.