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

Review · November 20, 2026

Audience-Bound Client Authorization for Token Exchange and Introspection

draft-zehavi-oauth-aud-bound-authz -00View on Datatracker →

This is a renamed and widened version of Audience-Bound Token Exchange, which we reviewed in October. The mechanism is the same. The new part is that it now applies to token introspection too, and invites future specs to apply it elsewhere.

Datatracker doesn't link the two. The old draft is still listed as active, and this one's history says "Initial version."

The problem it's solving

A resource server receives a token audienced to itself. It then asks an authorization server to do something with that token: exchange it for a downstream token, or introspect it.

How does the AS know the client asking is really that resource server?

Token Exchange leaves this to policy. may_act only works if the issuer knows the actor's client ID in advance. Token Introspection says the AS must check whether the token can be used at the calling resource server, and doesn't say how the caller maps to a resource.

The draft's answer: the token's aud names the resource, and the AS works out which client represents that resource.

How it fits in the working group

It's in charter. It extends RFC 8693 and RFC 7662 with processing rules and adds one parameter to Protected Resource Metadata. It defines no new endpoint, grant, or claim. It leans on Client ID Metadata Documents for client IDs that work across authorization servers.

The new relationship is that the resource server tells the AS which clients are allowed to act as it.

It still cites none of the adopted work that has the same question inside it: Identity Chaining, Transaction Tokens, and ID-JAG.

What the mechanism actually does

The AS picks exactly one audience value from the token. Then it establishes an "audience-client relationship" in one or more of three ways:

  1. Local policy. A configured mapping such as api://orders -> rs-orders-client.
  2. private_key_jwt validated with the resource's keys. The AS fetches the resource's metadata, follows its jwks_uri, and validates the client's authentication assertion with those keys.
  3. A new resource_clients metadata parameter. The resource publishes {authorization_server, client_id} pairs. Leave out authorization_server and the client_id is a CIMD URL.

For token exchange, a failure means an error response. For introspection, a failure means {"active": false}, the same answer a dead token gets.

Everything else stays in AS policy: which downstream audience is allowed, which scopes, what introspection discloses. The metadata says who represents the resource. It doesn't say what they may do.

What changed from the earlier draft

  • token_exchange_clients is now resource_clients.
  • Introspection is a second application, and Section 12 tells future specs how to add more.
  • The JWT in mechanism 2 is now always a client assertion. The earlier draft also allowed an actor_token and never said what was in it. That ambiguity is gone.
  • A new security section says introspection and exchange carry different risk, and that the relationship alone shouldn't permit exchange.
  • The unused references are cleaned up.

Introspection is the weakest case for discovery

The motivation for discovery is the multi-AS case. A resource server calls a foreign AS that has never heard of it, and the token issuer couldn't have put the right client ID in may_act.

Introspection never looks like that. The resource server introspects at the AS that issued the token. That AS chose the audience. It already knows the resource, and it already registered whatever credential the resource uses to call the introspection endpoint.

So for introspection, mechanism 1 is what every AS already does, and mechanisms 2 and 3 have the AS fetching metadata to learn something it configured itself.

The generalization also changes what resource_clients means. A client listed so it can introspect is now also a candidate for token exchange. The draft handles this with a SHOULD telling the AS to keep per-operation policy. That puts the config table back at the AS, which is what the metadata was meant to replace.

Significant interoperability concerns

The precondition still undercuts the discovery. The AS must first decide, by local policy, that the audience is a resource it trusts. Only then may it fetch metadata. An AS that keeps that list can add a client_id column to it. The draft still doesn't describe the case where discovery pays for itself, such as a resource team rotating client registrations without involving each AS admin.

The jwks_uri mechanism has an ordering problem. Step 1 in both processing lists is "authenticate the client." With mechanism 2, the keys that authenticate the client come from the token's audience, which the AS only learns in a later step. For introspection that means looking up a token for a caller that hasn't been authenticated yet.

It doesn't say which client ID the assertion carries. RFC 7523 requires iss and sub to be the client ID. Is that client registered at the AS? If it is, the AS already has its keys and a local mapping. If it isn't, any key in the resource's JWK Set can assert any client ID. The spec needs to pick one.

It still repurposes a key. RFC 9728 describes jwks_uri as the resource's keys, such as the ones used to sign resource responses. A resource publishing it for that reason today would be publishing client authentication keys under this draft. Section 13.4 says to use keys "only for purposes for which they are intended." It defines no use value or other marker that would let an AS tell the two apart.

Carried over and not fixed:

  • Section 8.2 says the AS SHOULD check the assertion's client ID against resource_clients. Section 9.4 says MUST. Same check.
  • "Bound to the globally scoped client identifier" is still undefined.
  • may_act "takes precedence," and the AS MAY also require both or neither. Nothing says what happens when may_act names a different client.
  • The AS MAY use any mechanism and MAY require several. Nothing in AS metadata tells a resource server which ones a given AS accepts.
  • The multi-AS examples still have a foreign AS accepting an access token issued by someone else as the subject token. Identity Chaining and ID-JAG avoid that on purpose.

Mechanisms that could be removed

Drop the jwks_uri mechanism. A CIMD client ID in resource_clients gives the same result, with keys that are unambiguously for client authentication.

Drop local policy as a named mechanism. One sentence saying the AS can always use its own configuration is enough.

Drop Section 12. A SHOULD-level template for future specs isn't needed, and the list of operations that act on a token on behalf of its audience is short. Revocation, for example, belongs to the client the token was issued to.

Reconsider the introspection application. Without a multi-AS story it adds text and no new capability.

What's left is the audience-selection rule and one metadata parameter for token exchange.

Writing quality

The prose is clean and uses BCP 14 keywords correctly. The metadata URL construction rules and the SSRF defenses are concrete.

It's 28 pages for one metadata parameter. The three mechanisms are listed in the abstract, the introduction, Section 5 and Section 7. The statement that the relationship "does not by itself authorize" anything appears in Sections 5, 7, 10.1, 11.1 and 13.1. Sections 8.1 and 13.3 repeat the same metadata check almost word for word.

Small things: the error codes are attributed to RFC 8693 when invalid_grant comes from RFC 6749, and the acknowledgments say "the authors" on a single-author draft.

Background of the author

Yaron Zehavi, Raiffeisen Bank International. His Datatracker profile shows zero published RFCs and six active drafts: this one, its predecessor, authority metadata, authorization request delegation chains, authorization schemas, and RAR metadata and remediation, which the working group adopted. Three more have expired.

He's not a new author. He's one of the most active individual contributors in the OAuth WG right now.

So what should happen next?

The first housekeeping step is to mark the old draft as replaced, so the two don't get reviewed in parallel.

The October feedback mostly still applies. The revision made the document broader and didn't make it smaller. The question for the list is the same one: does this stand alone, or is resource_clients a section in the Identity Chaining token exchange step? The jwks_uri mechanism should go either way.