The Scope Creep Automated reviews of new AI-related working group submissions

Review · September 2, 2026

AuthZEN Profile for Authorization Claims in JWT Access Tokens

draft-gazitt-oauth-authzen-claims -01View on Datatracker →

RFC 9068 registers three claims for JWT access tokens: groups, roles, and entitlements. It says what they're named and recommends drawing them from the SCIM user schema. It says nothing about where the authorization server gets the values.

In practice that's a directory lookup, a database join, or a vendor-specific extension point. This draft's observation is that "which groups is this subject in" is an authorization question, and it's currently being answered by something that isn't the authorization system.

The proposal: answer it with the OpenID AuthZEN Resource Search API. Give AuthZEN a subject, an action named member, and a resource type group, and the resources it returns are the groups the subject belongs to. The draft defines the bindings that make that mapping predictable and the rules for turning a search result into a claim value.

How this fits with the working group's existing work

Nothing changes on the OAuth wire. There are no new request parameters, no new token response fields, no new client behavior, and no new AS metadata. Every normative requirement in this document is a rule about what the authorization server does internally between receiving a token request and minting a token.

That's the first thing reviewers will want to think about. OAuth core deliberately leaves the AS's internal decision-making out of scope, and that's been a durable design choice. This draft argues that one specific piece of that internal behavior, populating enumeration claims, is worth profiling because the claim names are already registered and interoperable while their sources are not.

Dependency-wise it's light on OAuth RFCs and heavy elsewhere. RFC 6749 and RFC 9068 for the claims, RFC 7643 for the SCIM schema those claims came from, RFC 7519 for the claims registry. The two things doing the real work are both outside the IETF's control: the OpenID Foundation's AuthZEN Authorization API 1.0, and draft-gazitt-oauth-authzen-issuance, a companion individual draft from the same author.

This is the third of three related drafts. The other two are the issuance profile, which externalizes the decision to issue a token at all, and a token exchange binding. Reviewing this one in isolation gives an incomplete picture.

The independence claim doesn't quite hold up

Section 1.2 and Section 3 both insist this document is independent of the issuance profile. An AS that keeps its own issuance decision can still externalize claim enrichment.

Then Section 6.1 says the search subject MUST be the same entity as "the subject of the gate tuple of [ISSUANCE]." Section 2 imports "gate tuple" and "scope tuple" as terminology from ISSUANCE. And Section 15.1 registers three entity types into a registry that ISSUANCE establishes.

So the document can't actually stand alone. If the issuance profile never advances, this one has terminology it can't resolve and an IANA section with nowhere to register. Either the normative dependency should be acknowledged and the standalone framing dropped, or the subject requirement and entity type registrations need to be restated in terms of RFC 9068 and RFC 6749 alone. The second is probably possible and would make the document genuinely independent.

Interoperability

The only thing here that two independently built products need to agree on is the binding table. Three rows: groups/group/member, roles/role/assignee, entitlements/entitlement/holder. An AS and a PDP from different vendors, neither configured for the other, only interoperate because both arrive at group plus member without being told. The draft says this well in Section 5.3.

The claim value shape is the real problem

Section 8.1 renders each claim as a flat JSON array of strings. RFC 9068 recommends the SCIM complex form, where each value is an object with a value sub-attribute plus optional $ref, display, type, and primary. An AuthZEN search result only carries an id, so the complex form would be an array of single-key objects on every request the token is presented with. The reasoning for flattening is sound.

The problem is what the section says next. An AS that needs the complex form for a resource server that requires it isn't prevented from producing it, and "a resource server cannot infer which form to expect from the presence of this profile."

That's the one piece of this actually visible to a second implementation, and it's left unresolved. A resource server parsing groups still has to handle both shapes, exactly as it does today. Whatever else this profile fixes on the AS side, it doesn't fix that. If the working group takes this up, pinning the array-of-strings form as a MUST for anything claiming conformance, and saying explicitly that RFC 9068's SHOULD is being overridden, would be worth doing.

Deployment overrides undercut the registry

Section 5.3 lets a deployment configure action names other than the registered ones, and Section 6.3's audience context key is explicitly advisory. A PDP MAY use it to scope results and MAY ignore it, and the AS MUST NOT assume anything either way.

The draft is honest about both. It calls the registered bindings "defaults, not constraints" and says a deployment needing real target scoping should register a binding whose resource type distinguishes the target. That's the right escape hatch, but it means the same subject at two conforming deployments can get differently scoped groups claims with no way for the resource server to tell.

No default failure behavior

Section 10 gets the important part right. A failed search MUST NOT render the claim and MUST NOT substitute an empty array, because an empty array asserts the subject holds nothing. Section 8.4's rule that a successful empty search MUST render [] rather than omitting the claim is a good call and correctly justified: presence means the enumeration happened.

But whether the AS then issues a token without the claim or fails the request is left entirely to deployment configuration, with no default specified. The draft's reasoning is that only the operator knows whether a resource server treats the claim's absence as a value. Fair, but combined with the shape ambiguity above, a resource server has no way to distinguish "you hold nothing," "the lookup failed," and "this deployment doesn't bind this claim" beyond guessing from the claim's presence.

Things that could come out

Section 9 on pagination mostly restates AuthZEN Section 8.2. The one new normative bit is the SHOULD to bound page-following, which could be two sentences.

Section 11 registers no discovery capability and spends a page explaining why not, then describes an optional deployment-time check using the Action Search API that "nothing in this document requires." That's design rationale, not specification.

Section 12's error mapping table restates conditions already stated normatively in Sections 8.2, 9, and 10. The only new content is the HTTP 500 guidance, which belongs in Section 10.

Section 14.3 is a genuinely interesting argument about why the draft defines no provenance or freshness claim. "Trusting an issuer does not make it a witness to its own freshness" is a good line. It's also two pages of philosophy defending a decision not to add something, which reads better as a mailing list post than as an RFC section.

The new IANA registry in Section 15.2 is worth questioning. It has three entries, requires the claim name to already be in the JWT Claims registry, and requires the resource type to be in another new registry created in the same document. For a mechanism this narrow, a table in the RFC plus a note about how to extend it may be enough.

Writing quality

The prose is unusually good. Sentences are well formed, the reasoning is careful, and the author clearly anticipated objections. Section 7 on ordering is a model of explaining why two requirements that look contradictory aren't: search results are semantically after the decision, but MAY be issued concurrently for latency, and the two consequences for operators (wasted work on denials, PDP search volume not tracking tokens issued) are spelled out.

The problem is that it reads as an essay about a specification rather than as a specification. Twenty-four pages carry maybe six pages of normative content. There's an "Editor's note" inviting feedback on the assignee and holder names, an "Implementer's note" on failure defaults, and repeated meta-commentary: "The choice is observable, so it has to be made deliberately," "this document does not pretend otherwise," "The requirement is not merely for tidiness." Section 1.1 exists to explain why a search rather than an evaluation, which is the kind of thing that belongs in an appendix or a design rationale document.

There's also a formatting habit worth flagging. Several normative statements are set in emphasis, like Section 7's "An AS MUST NOT include an enrichment claim in a token it does not issue" and Section 10's "MUST NOT". RFC style doesn't do that, and the RFC Editor will strip it.

The editor's note about assignee and holder is the right instinct. A single action name with the resource type carrying the distinction would be simpler, and the argument for distinct names ("policy is easier to read when names differ where the concepts do") is a preference rather than an interoperability requirement.

Author background

Omri Gazitt, listed as Independent. His Datatracker profile shows zero published RFCs and three active Internet-Drafts, all part of this AuthZEN series. New to IETF authorship, though the acknowledgments point to involvement in the OpenID AuthZEN interoperability events, including a December 2025 identity provider scenario that demonstrated exactly this mechanism.

The implementation experience shows. This draft is written by someone who has actually made searches populate token claims and hit the edge cases: transitive search returning duplicates, pagination without an atomic snapshot, what to do when the enrichment dependency is down.

What changed in -01

Two things. Section 14.3 on provenance and freshness is new. And the search context's carve-out for the enforcement point capability declaration came out, along with the declaration itself from the example, because the issuance profile no longer defines it.

Where feedback would be most useful

The threshold question is whether the OAuth working group is the right venue. The normative dependencies run toward the OpenID Foundation, nothing on the OAuth wire changes, and the mechanism is a rule for how an AS talks to a PDP. The counterargument is that the claims being populated are RFC 9068 claims and the binding registry needs an authoritative home.

Beyond that, three specific things:

  • Resolve the independence question. Either drop the standalone framing or restate the subject requirement and entity types without reference to the issuance profile.

  • Pin the claim value shape. The array-of-strings decision is right, but leaving the complex form as a permitted alternative means resource servers gain nothing.

  • Cut the document roughly in half. The rationale is good, and it can live in an appendix or on the mailing list.

For context on the companion work, see draft-gazitt-oauth-authzen-issuance, which externalizes the decision to issue a token at all and is the larger of the two proposals.