Review · November 20, 2026
OAuth 2.0 Token Exchange with Actor-Bound Tokens
The Problem
Token Exchange gives the token to whoever sent the request. With DPoP or mTLS, that token is bound to the requesting client's key.
Agent systems often want one party to get a token for another. An orchestrator approves an action and fetches a token for the sub-agent that will perform it. Today the orchestrator either hands over a token bound to its own key, which the sub-agent can't use, or a bearer token, which anyone can use.
This draft lets the requesting client ask for a token bound to the actor's key instead. The actor has to sign a proof agreeing to it.
How It Fits In
This is a small extension to Token Exchange (RFC 8693). It fits the OAuth working group charter.
It adds one mechanism and no new roles or endpoints. The new relationship is between the requesting client and the actor: the client receives a token it cannot use and has to deliver it.
It depends normatively on Token Exchange, DPoP, mTLS, RFC 7800 (cnf), RAR, JCS, and AS Metadata. It also cites the OAuth 2.1 draft normatively, only for the token error response.
The informative references cover six individual drafts, including transaction challenge, transactional access tokens, client instance identifiers, and the actor profile. The draft positions itself as the single-AS version of the requested_cnf idea in Identity Chaining. That text is an example in an appendix of the chaining draft, not a defined mechanism.
Summary of the Mechanism
There are three parties: the requesting client (the approver), the actor (the agent that will act), and the authorization server.
- The actor creates an actor proof. This is a JWT signed with the actor's key, with
typofactor-proof+jwt. It carriesaud(the AS issuer),iat,exp,jti, andath, a hash of the actor token. - The actor gives the proof and its actor token to the requesting client, for example in the message asking for approval. The actor never talks to the AS.
- The requesting client sends a normal token exchange request with
subject_token,actor_token, and the newactor_proofparameter. - The AS checks that the actor token has a
cnfclaim and that the proof is signed by that key. It rejects ajtiit has seen before. - The AS issues a token with
cnfset to the actor's key andactnaming the actor. The response includesactor_bound: true. - The requesting client hands the token to the actor. The actor uses it with its own DPoP proof or client certificate. The resource server needs no changes.
The actor can also limit what the proof is used for. An optional azh claim holds a hash of the authorization_details the actor asked for. The AS must issue details equal to or narrower than those. If the approver narrowed the request, it sends the actor's original details in a new actor_authorization_details parameter so the AS can check the hash.
The draft also adds an invalid_actor_proof error and an actor_bound_token_exchange_supported metadata flag.
Interoperability
The core flow is well specified. Two implementers reading sections 2.1 to 2.3 would build the same thing for the DPoP case with exact-match authorization details.
Significant interoperability concerns
"Narrower" is not defined here. The AS must reject any authorization details type whose narrowing rules it does not know. Those rules come from draft-nandakumar-oauth-rar-narrowing, an informative reference to another individual draft by the same author. In practice only exact equality will interoperate until that draft exists and types adopt it.
azh only covers authorization_details. An actor that signs azh has not limited scope, resource, or audience. The requesting client can still pick those freely. The security considerations say azh "limits that request to the details it asked for", which overstates what it does for deployments that use scopes.
The proof is not bound to the requesting client or the subject token. The draft admits this. Anyone who gets hold of the proof and actor token before they are used can choose which exchange they go into. An optional claim naming the intended requesting client would close this cheaply.
Nonces don't work. The proof has an optional nonce claim "as for DPoP", but the actor never contacts the AS. The draft says delivery of the nonce is out of scope. An AS that requires a nonce cannot interoperate with any actor.
Refresh tokens are not mentioned. If the AS returns a refresh token, the draft doesn't say who it is bound to or who can use it. The requesting client holds it, and the actor holds the key.
Trust in the actor token is left to RFC 8693. One of the two motivating cases is a sub-agent "that has no relationship with the authorization server". The AS still has to validate that sub-agent's actor token and trust its issuer. The draft doesn't say how.
Long-lived proofs for human approval. The draft allows a longer exp when a person has to approve. The AS stores every jti until exp, and the unbound proof stays usable for that whole window.
Mechanisms that could be removed
The CIBA paragraph. Section 4 says the same parameters "can be carried" in a CIBA request. That is a one-sentence extension to an OpenID Foundation spec with no processing rules. It should be its own document or come out.
The ID-JAG case. The draft says an assertion like an ID-JAG can carry the cnf and act claims, and that the party accepting the grant "requires proof of possession". It doesn't say how that proof is presented. This needs a real section or should defer to the ID-JAG draft.
mTLS support. The x5c header and certificate-bound output add a second code path. The agent use cases in the introduction all point to DPoP. Dropping mTLS from the first version would shrink the test matrix.
The nonce claim. It can't be populated in the main flow the draft describes.
Writing Quality
The writing is good. It uses standard IETF language, sentences are clean, and normative keywords are used carefully. The security considerations are honest about what the proof does not cover.
It is short: about seven pages of content. That is mostly a strength. The mechanism is easy to review in one sitting.
It is too short in a few places. There is no decoded example of an actor proof and no example token response. "Relationship to Other Work" mixes comparisons with other drafts and what are effectively new extension points for CIBA and ID-JAG. The Acknowledgments section contains a design statement, not acknowledgments.
Reusing the ath claim name from DPoP for a hash of the actor token is reasonable, but it is a different token than DPoP hashes. The draft should say so in the claim definition, not only in the IANA section.
Authors
Suhas Nandakumar (Cisco) has 9 published RFCs from 2013 to 2025, all in real-time media: STUN and TURN URI schemes, RTP taxonomy, SDP and ICE. He has 36 active Internet-Drafts. Most are Media over QUIC. About ten are recent OAuth, WIMSE, and AGENTPROTO drafts, including draft-nandakumar-oauth-delegation-revocation and the PACE agent protocol drafts. He is an experienced IETF author who is new to OAuth and writing at high volume right now.
Cullen Fluffy Jennings (Cisco) has 29 published RFCs from 2002 to 2026, including SIP identity (RFC 3325, RFC 4474), NAT behavior (RFC 4787), and JSEP for WebRTC (RFC 8829). He has 30 active Internet-Drafts, with the same split: mostly Media over QUIC plus the same set of agent, OAuth, and WIMSE drafts co-authored with Nandakumar. He is a long-time IETF contributor, also new to the OAuth working group.
This draft is one piece of a larger set from the same two authors. It cites their RAR narrowing draft and is designed to plug into their agent approval flow.
Bottom Line
The core idea is sound and small: let a token exchange client ask for a token bound to someone else's key, with that party's signed consent. It turns the requested_cnf example from Identity Chaining into a defined mechanism for a single AS.
So what should change before the working group spends time on it? Bind the proof to more of the request (scope, resource, and the requesting client), decide what happens with nonces and refresh tokens, and cut the CIBA and ID-JAG sketches. The narrowing rules need to land somewhere normative, or the draft should require exact match for now.