AI agent authentication vs authorization
By Nodrail · 4 min read · Published
On this page
Authentication establishes which identity a system accepts for a caller. Authorization decides whether that caller may perform a particular operation on a resource. A successfully authenticated AI agent can still be forbidden from deleting a ticket or reading another customer's records.
The distinction is established in OWASP's authorization guidance. For agent workflows, it helps to separate the installation, any person being represented, and the person reviewing an action.
The examples below describe a design, not a specific identity provider integration or a test result.
Give identity and permission different jobs
Suppose a team connects a logical agent called support-agent. It has one installation for development and another for production. Each installation can present its own credential to the policy service.
A request authenticated as the production installation establishes that credential's context. The authorization step still has to decide whether the proposed tool, operation, target and effective arguments are in scope.
| Question | Evidence needed in this example |
|---|---|
| Which logical agent is this? | A stable agent record linked to the authenticated installation. |
| Which installation is calling? | A valid credential bound to that installation and workspace. |
| May it act on this ticket? | A current rule and trusted resource context. |
| Is it representing a person? | A verified delegation relationship, when that mode applies. |
| Who can approve this request? | A reviewer with current authority over this action. |
An identifier in a request body is a claim, not proof. Derive the authenticated context from the credential and verify any additional relationship before using it to grant access.
Understand what a credential demonstrates
A bearer credential can be used by whoever possesses it. RFC 6750 defines that property explicitly. A stolen credential can therefore make a different process appear to be the installation associated with it.
For our design, use a distinct credential for each installation that must be independently revocable. Rotation changes the credential without erasing the logical agent's history. A run identifier can correlate one task's events; it does not by itself establish another security identity.
Also distinguish the credential for the governance service from credentials for the ticket provider. Authenticating to the former does not remove the latter from the customer runtime. If the agent can use the provider directly, it may take a route outside the policy integration.
Keep the identity constant while decisions change
Here is an illustrative policy for one authenticated production installation:
| Proposed request | Policy result | Why |
|---|---|---|
| Read ticket SUP-42 in Support | Allow | A scoped read rule covers that project and operation. |
| Send a specific reply to SUP-42 | Require approval | This externally visible write is eligible for review. |
| Delete SUP-42 | Deny | An explicit rule prohibits deletion. |
| Read FIN-9 in Finance | Deny | That resource is outside this agent's permitted project. |
The credential has not changed between rows. The requested action and its policy match have.
Decide which facts the system can trust. A statement such as “this ticket belongs to Support” supplied by the agent is not equivalent to an authoritative resource lookup. If a rule only compares declared parameters, describe it that way. Missing context should not quietly expand access.
Treat delegated user authority as another relationship
An agent may operate for a team account or represent a particular user. Those are distinct modes. Assigning a teammate to manage the agent does not automatically prove that every call acts on that person's behalf.
For a user-representing operation, our example needs four decisions: authenticate the acting installation; verify the represented person and delegation; constrain the operation by the agent's policy; and obtain any required human review. An approval is additional authority for its defined scope, not proof that a delegation existed in the first place.
RFC 8693 distinguishes delegation, where actor and represented subject remain identifiable, from impersonation semantics. It is one protocol reference; adopting a token format does not determine your business permissions or mean that a particular product implements token exchange.
When delegation is unavailable or unverified, do not label installation-level activity as user-authorized. Retain what the evidence actually establishes.
Authenticate the reviewer and authorize the decision
A signed-in teammate may have permission to view a request without permission to approve it. A runtime credential should not inherit human review capabilities merely because it can submit an action.
Bind review to the operation that will be attempted: caller, installation, target, material parameters and expiry. If the request changes, require a new decision. The OWASP agent security guidance recommends exact-action integrity controls around consequential operations.
For our policy, a human review cannot overrule the explicit deletion prohibition. A standing rule is a separate policy change. See the human approval workflow guide for the lifecycle of a pending request.
Record the identities without claiming an unobserved effect
An investigation should be able to connect an installation, action ID, policy version and reviewer. Include a represented user only when that relationship was verified. Keep a client-reported outcome separate from the permission decision and from any independent provider evidence.
Before adopting an integration, ask it to demonstrate wrong-workspace rejection, revoked credentials, unauthorized reviewers and changed arguments. These are validation requirements for this design, not executed tests reported here.
Authenticate the caller in your integration, then map that identity and the requested operation into the policy evaluation. Inspect the descriptor in the library guide and review the security model.