AuthZEN
OpenID AuthZEN · Authorization API 1.0 · AuthZEN Authorization API
By Boston Identity · Explained · Updated · 5 min read
The short answer
AuthZEN is the OpenID Foundation's standard API for asking a separate policy service "may this user do this to that?" and getting back a yes or no, so applications stop hard-coding their own access rules.
- What it is
- A standard
- Defined by
- OpenID Foundation, Authorization API 1.0 (Final)
- Who it's for
- Teams moving access rules out of apps
- In our briefings
- 120by month, last 12
What is AuthZEN?
Authorization API 1.0 is the specification from the OpenID Foundation's AuthZEN working group, published as Final on 11 January 2026.S1 It defines how a Policy Decision Point (PDP), the service that decides, and a Policy Enforcement Point (PEP), the application that asks and then acts, talk to each other without knowing each other's inner workings.S1 Abstract The point is that one vendor's application can ask another vendor's decision service, with neither tied to a single implementation.S1 §1
The spec fixes the shape of the question and the answer, not the rules behind them. How a PDP writes and stores its policy is out of scope.S1 §2 The core call asks one question, such as whether Alice can view document 123.S1 §3 A batch call asks several at once, and three search calls ask the reverse: which documents Alice can view, or who can view document 123.S1 §3
AuthZEN is not a policy language or a policy engine. It is the interface in front of one, and it says nothing about how the engine decides.S1 §2
How does AuthZEN work?
Two actors take part: the application that needs an answer (the PEP) and the decision service that gives it (the PDP). The walk-through follows a single access question over the HTTPS binding, the one transport every compliant PDP must support.S1 §10
Figure steps
- The application finds the decision service
- The application builds the question
- The application sends it to the decision service
- The decision service answers yes or no
- The application enforces the answer
The application asks the decision service one subject-action-resource question over HTTPS and must stop on a false answer; the policy behind the answer stays inside the decision service.
Drawn from Authorization API 1.0, sections 6, 9 and 10
Swipe for the full figure →
The application finds the decision service
A PDP is recommended to publish metadata about itself at a well-known address, formed by inserting /.well-known/authzen-configuration between the host and any path.S1 §9, §9.2 The metadata must name the PDP and the URL of its single-evaluation endpoint, and the other endpoints are optional.S1 §9.1.1 The application must discard the metadata if the PDP identifier inside it does not match the address it fetched from.S1 §9.2.3
The application builds the question
A single access question carries a subject (the user or machine the question is about), an action (what is to be done) and a resource (what it is done to), and may add a context such as the time of day.S1 §6.1 Subjects and resources each need a type and an id, and any of them can carry extra properties, such as a department or an IP address.S1 §5.1, §5.2
The application sends it to the decision service
The application posts the request as JSON (JavaScript Object Notation) over HTTPS, to the endpoint the metadata names or, failing that, the default path /access/v1/evaluation.S1 §10.1 The connection must be secured, for example with TLS (Transport Layer Security).S1 §11.1 The PDP should authenticate the caller, and the spec leaves the method open, naming mutual TLS, OAuth or an API key as options.S1 §11.2 The spec recommends support for OAuth 2.0.S1 §2
The decision service answers yes or no
The PDP returns a decision, a required true or false, and may add a context object carrying reasons or hints such as a request for stronger login.S1 A deny comes back as an ordinary success with "decision": false, while HTTP error codes such as 401 mean the request itself failed.S1 §10.1.2
Why does AuthZEN matter now?
Ping Identity has already committed to AuthZEN in PingOne Authorize, and on that and other signs we forecast that by March 2028 at least two of Okta, Microsoft Entra and Ping will sell per-action authorization for AI agents as a generally available product.1 Raut argues an enterprise has to assemble agent controls from several pieces, with OpenID AuthZEN supplying the policy decision.2
The working group is also building on top of the API. On 15 June 2026 it announced two approved working-group drafts: one for policy that answers "not yet, here is what is required", such as a manager's approval, and one that maps other request formats, starting with tools in the Model Context Protocol (MCP), onto AuthZEN's subject, action, resource and context.3 McGuinness cautions that AuthZEN cannot help with agent intent until there is a shared vocabulary of agent purposes, and that work has not been done.4
Where does it fit?
| Compared with | How AuthZEN differs |
|---|---|
| XACML | AuthZEN uses the PDP and PEP split that XACML defines.S1 §1 It standardizes the request, response and discovery between the two, and leaves the PDP's policy language out of scope.S1 §2, §9 |
| Open Policy Agent | We read Open Policy Agent as something that could sit on the PDP side of the call. AuthZEN leaves a PDP's policy language, architecture and state management out of scope,S1 §2 so it governs the question and answer, not how the engine reaches its answer. |
What goes wrong?
- The decision service trusts the application. The spec says the model "assumes the PDP must trust the PEP",S1 §11.4 so a subject, property or context value the application sends is taken as true.5
- Caller authentication is only a should. The PDP should authenticate the application, and the method is out of scope.S1 §11.2 Without it, an attacker can flood the PDP or probe it with many requests to learn its policies.S1 §11.2
- There is no way to say an agent acts for a person. A subject is one user or machine, so "agent X for user Y" has to be packed into properties or context keys the standard does not name, and two PDPs need not read them the same way.5
- "AuthZEN support" can mean little. Only the single-evaluation endpoint is required, and the meaning of the context in a decision is left to each implementation.S1 The spec names no conformance test.6 The working group announced vendor interoperability results in May 2024, before the final spec,S2 and it says it is developing a conformance certification program.S3
- TLS covers one hop. It protects one point-to-point hop, so a PDP may sign its decisions to let the application check them across gateways.S1 §11.6
- Search results are not a promise. A subject, resource or action a search returns should come back true when evaluated, but the spec says that outcome is not guaranteed.S1 §8.1
What should a security leader ask?
- 01
Which of our applications send AuthZEN requests, and how does each one prove its identity to the decision service?
- 02
When an agent acts for a person, which property or context keys carry that, and does every decision service we use read them the same way?
- 03
Does our vendor's AuthZEN support go beyond the single-evaluation endpoint, what does its decision context contain, and what interoperability results or certification can it show for the final spec?
- 04
Which applications treat a decision with context they do not understand as a deny?
- 05
If the decision service is down or slow, do our applications fail closed?
Sources
- S1
Authorization API 1.0OpenID Foundation, AuthZEN Working Group · Final, 11 January 2026
- S2
AuthZEN Work Group Announces Authorization Interop ResultsOpenID Foundation · Published 29 May 2024
- S3
AuthZEN Working Group overviewOpenID Foundation · Working group page, read 5 November 2026
›Evidence: 6 claims from 5 briefings
- 1BI-CALL-2026-0049 · When the model cannot be trusted, control what the agent can reach, checked on every actionBy March 2028, at least two of Okta, Microsoft Entra and Ping Identity will sell per-action authorization for AI agents, a policy decision evaluated at each tool or API call rather than at sign-in, as a generally available product. Mattson's 'complete separation of concerns', a security layer distinct from the agent layer, is his company's roadmap and would be discounted on its own; what makes the call is that Ping has already committed to AuthZEN in PingOne Authorize, Uber and Sierra built the layer themselves because nothing sold it, and WorkOS argues authorization has to live outside the gateway. One vendor or none by that date settles this wrong.
- 2BI-FIND-2026-0241 · Keeping an agent's delegation chain is optional, and room in the HTTP header decides whether it survivesNo standard covers the control an agent most needs: asking for more authority in the middle of a task. Raut's point is that an agent given a goal works out its own path and permissions as it goes, so least privilege set in advance does not hold. Step-up authentication has protocols; step-up authorization has almost none, and there is no single standard for what he calls least agency. An enterprise has to assemble it from SPIFFE for workload identity, transaction tokens for context, OpenID AuthZEN for the policy decision and an agent-control standard he says is coming and does not name.
- 3BI-FIND-2026-1013 · OpenID's AuthZEN group approves a draft that lets policy tell an agent 'not yet, and here is what is required'The OpenID Foundation's AuthZEN working group has approved two working-group drafts, announced on 15 June 2026. AARP defines common patterns for requesting, tracking, satisfying and re-evaluating what policy needs before it can authorize an action, such as approval, consent, delegated authority or an attestation; in the post's example, an agent's payment above a threshold comes back flagged as approvable along with what is needed, and it goes ahead only after a manager approves and policy is evaluated again. The foundation likens AARP to CIBA: an asynchronous, out-of-band step that a person or an automated governance system can satisfy. COAZ maps other information models into AuthZEN's subject-action-resource-context request so gateways and other enforcement points can call a compatible decision service, and its first target is letting MCP tools expose the authorization checks a call needs.
- 4BI-FIND-2026-0206 · A user's approved request is not yet authority an agent can be held to, and most deployments skip that stepNo shared vocabulary of agent purposes exists, so every implementation invents its own, and McGuinness says AuthZEN, the OpenID Foundation's externalized-authorization standard, needs that upstream work done before it can help. The harder part is the approver: a human signing off on a compiled authority model cannot verify what it grants, a gap he compares to the difference between informed consent and a terms-of-service checkbox, and one that widens as template libraries grow. One author's diagnosis, uncorroborated, and the reason the intent question stays open on our agenda.
- 5BI-FIND-2026-0867 · OpenID's AuthZEN API can't say an agent acts for someone, and its policy service trusts whatever the app sendsAuthZEN Authorization API 1.0 has no way to say that an agent is acting for a person. The words agent, delegation and on-behalf-of do not appear; a subject is one user or machine with a type, an id and free-form properties, so an application can only encode "agent X for user Y" in property or context keys the standard does not name, and two decision services need not read them the same way. Whatever the application sends is taken on trust: the spec says "the PDP must trust the PEP" (the policy service trusts the enforcing application), leaves authenticating the caller out of scope, and only says the service SHOULD authenticate it. A delegation claim passed through AuthZEN is the calling application's word, with no token or chain to check.
- 6BI-FIND-2026-0868 · OpenID's AuthZEN API can't say an agent acts for someone, and its policy service trusts whatever the app sends"AuthZEN support" can mean very little. A compliant decision service must implement the HTTPS and JSON binding. Its discovery metadata requires the single-evaluation endpoint and makes the batch-evaluation and three search endpoint URLs optional; publishing that metadata is only recommended. A decision is a required yes or no plus an optional context whose format is left to each implementation; the spec only suggests borrowing keys such as HTTP status codes, so reasons, obligations and step-up hints are not guaranteed to mean the same thing across products. The text names no conformance test or implementation, so a vendor's claim of support guarantees little beyond the HTTPS binding and the single evaluation call.
›In our briefings
The claim each of the newest briefings on AuthZEN would have you leave with.
A person's approved request is not, by itself, authority an agent can be held to, McGuinness argues, and relatively few deployments perform the step that would make it one. That step converts approved intent into a bounded authority artifact an enforcement point can evaluate. By his account, many systems instead fall into one of five patterns: scope enumeration, app-local mapping, LLM-generated requests, broad standing authority with prompt discipline, or a formal policy language. He says the most common land in the third or fourth, and that where nothing makes the conversion, the agent's own judgment becomes the effective policy engine. This is one author's argument, but two production designs we have read fit it. Uber's agent tokens carry the full chain of who asked through every hop, and Sierra's gateway runs interactive agent work as the user, with that person's permissions. In our reading, neither records what the person approved.
A user's approved request is not yet authority an agent can be held to, and most deployments skip that stepBI-FIND-2026-0202
Much of the new text in -01 says how an agent's access ends, in plain prose with no MUST or SHOULD. Section 8 now says a posture check at provisioning "establishes trust at the moment a credential is issued or renewed" and does not by itself show that posture still holds at a later request. It offers two existing remedies: shorter credential lifetimes, so posture is re-assessed at each renewal, or the event mechanisms of Section 11, so a posture change after issuance revokes or down-scopes the agent's credentials and tokens. Deprovisioning becomes a third phase after provisioning and rotation. At a minimum renewal stops and the short-lived credentials lapse; where the change must take effect before expiry, the issuer signals relying parties and the issuers of any derived credentials. A new Section 9.3 bounds a long-lived session or stream by the credential that opened it: the accepting party re-validates when that credential expires, and both sides end or down-scope the session on a revocation or risk signal. A retired agent's identifier must not be reassigned to a different agent.
WIMSE's agent draft adds how an agent's access ends and where policy engines fit, without a new requirementBI-FIND-2026-2029
Four voices say a point-in-time access review cannot keep pace with how often access changes. Rick Wagner of SailPoint says periodic reviews judge access at one moment, invite rubber stamping and will not scale to agents, and he wants real-time authorization instead. Omar Khawaja of Databricks says a review that snapshots one day fails when access changes every week, and he wants versioning added to identity. Justin Devine, an RSM consultant, says no team can staff human reviews over every identity a cloud estate creates. Andrej Safundzic of Lumos says Lumos runs a daily review agent at a couple of customers in place of the quarterly certification. Their fixes differ. Ciena's Steve Rennick, on the same show as Devine, wants runtime policy to end campaign certifications, and the exchange notes that runtime policy needs the same clean entitlement data that roles did.
A point-in-time access review can't keep pace with changing access, four voices say, and four vendors now put AI inside the reviewBI-FIND-2026-1747
›Cited in 12 briefings
Every briefing that discusses this term, newest first.
- A user's approved request is not yet authority an agent can be held to, and most deployments skip that step
Article · · Control Plane by Karl McGuinness · Karl McGuinness · Mar 17, 2026
- WIMSE's agent draft adds how an agent's access ends and where policy engines fit, without a new requirement
Draft · · IETF WIMSE Working Group · Pieter Kasselman et al. · Sep 15, 2026
- A point-in-time access review can't keep pace with changing access, four voices say, and four vendors now put AI inside the review
Field signals · · 9 voices from the field · Nov 2024 to Aug 2026
- OpenID's AuthZEN API can't say an agent acts for someone, and its policy service trusts whatever the app sends
Article · · OpenID Foundation · Omri Gazitt et al. · Jan 11, 2026
- Schrock's receipts prove a named key approved one exact agent action, and the draft lists at length what they cannot prove
Draft · · IETF Datatracker · Iman Schrock, EMILIA Protocol · Sep 12, 2026
- Before an agent pays or changes data, Schrock's last check treats it as untrusted and keeps the record itself
Draft · · IETF Datatracker · Iman Schrock, EMILIA Protocol · Sep 26, 2026
- Eight panels agree an agent acts for a person, never as that person. When agents hand work to other agents, nobody had an answer.
Conference · · Las Vegas · June 2026, a one-day summit track inside Identiverse 2026; the eight panels were published as podcast episodes between 18 August and 7 September 2026
- OAuth handles an agent in one company, working while you watch. The OpenID paper says that is where it stops.
Article · · OpenID Foundation · Tobin South et al. · Oct 2025
- Keeping an agent's delegation chain is optional, and room in the HTTP header decides whether it survives
Podcast · · The Non-Human & AI Identity Podcast · NHI & AI Identity Summit : The New NHI Risk Reality · Aug 26, 2026
- Ping moved agent identity from roadmap to shipping product, and let AI agents run its admin tools too
Conference · · Ping Identity · Austin · 2-day event (Day 1 / Day 2), 2026
- OpenID's AuthZEN group approves a draft that lets policy tell an agent 'not yet, and here is what is required'
Article · · OpenID Foundation · Serj Hallam · Jun 15, 2026
- A proposed standard for agent authority advises against delegating it across company lines by default
Paper · · arXiv · Partha Madhira · May 12, 2026
Related terms
What appears alongside AuthZEN in the same briefings.
Open Policy AgentARIA (Agent Relationship-Based Identity and Authorization)OpenID Foundation AI Identity Management Community GroupXACMLAI Agent Authentication and AuthorizationCAEPLumosAxiomatics
Was this explainer useful?
Share this explainer
Pass it on. Every claim has a permanent number, so what you share today still points at the same words next year.