Review · September 1, 2026
Handing the Subject's Signing Key to the Resource Server
The problem it's solving
A user signs something. Not the client, the user. A payment instruction, a delegation constraint, some application-layer statement that narrows what the client is allowed to do on their behalf. The client carries both the access token and that signed object to the resource server.
The resource server now needs the user's public key to check the signature. OAuth gives it no way to get one.
That's the gap, and it's a real one. DPoP and mTLS-bound tokens bind a token to a key held by the client presenting it. cnf says "whoever shows me this token should also prove they hold this key." None of that helps you verify something the human signed before the client ever got involved.
So this draft defines a subject_keys claim whose value is a JWK Set, carried either as a top-level claim in an RFC 9068 JWT access token or as a top-level member of an introspection response. The authorization server becomes a key directory for its own users, and the binding is scoped to the tuple of issuer, sub, audience, and token validity. Nothing more.
The draft is emphatic that this is not a grant. Evidence verified with one of these keys can narrow the token's authority. It can never widen it.
How it fits in the working group
Mechanically this is about as clean an extension as you can file. One JWT claim registration, one introspection response member registration, no new endpoint, no new grant type, no new request parameter, no change to any existing flow. It doesn't need an Updates: header and doesn't claim one.
Nine normative references, six informative. That's a lot of citations for one claim. Most of them are real dependencies, since the document genuinely builds on RFC 9068, RFC 7662, RFC 7517, and RFC 7638.
It also explicitly stays out of cnf's way, out of RAR's way, and out of sub_jwk's way, with a short section on each. Section 9.1 goes further and says using one key for both proof-of-possession and subject evidence is NOT RECOMMENDED. That's the right call and it's rare to see a new draft draw that line proactively.
This replaces draft-chu-oauth-as-attested-user-cert, which asked for an AS-signed user certificate via a RAR authorization details type. The certificate is gone, the RAR type is gone, the request is gone. That's a big simplification and it's the right direction. It also left a hole where the request used to be, which I'll come back to.
Here's the harder question for the chairs. The thing this key is for is out of scope. The signed object, its syntax, its replay protection, its binding to the request, how it gets to the client, all out of scope. Section 8.3 lists twelve things an evidence profile MUST define, and no such profile exists in the OAuth WG today. So this registers a claim with no consumer.
That's not automatically fatal. Registering a claim is cheap and the semantics are worth pinning down before three vendors invent three versions. But a 30-page standards-track document for a claim nobody can use yet is a different proposition than a registration request, and the WG should decide which one it's being asked for.
The other thing worth saying: the obvious consumer is right in front of the working group. "The human signed a constraint, the agent presents it, the API checks it" is the agent delegation problem. The draft mentions act exactly once, in passing, in Section 8.5, and never positions itself against the pile of agent delegation drafts filed in the last month. Doing that would help this document a lot more than another page of security considerations.
What the mechanism actually does
The AS establishes a binding between a user and a public key. How is out of scope, though Section 7.1 imposes eight requirements on it anyway.
When issuing a token, the AS looks up active keys bound to that user and authorized for that specific audience, checks them against the JWK Set rules, makes sure the token doesn't outlive the binding, and includes them. It must not copy a key the client supplied in the request. If there's no valid key it omits the claim or fails issuance, and it must never emit an empty JWK Set.
The RS validates the token normally, pulls subject_keys, applies local limits on size and key count, selects exactly one key by an integrity-protected kid from the evidence, and verifies. It must not try every key until one works.
Then Section 8.4, the part the authors clearly care most about:
Effective Authority =
Token Authority
INTERSECT Evidence Authority
INTERSECT Resource Server Local Policy
Three inputs, all of them limits. Evidence can never add a scope, a resource, an action, an amount, a recipient, a time period, or a delegation right that isn't already in the token. The access token doesn't get to override a narrower constraint in the evidence either. When the two express constraints in different vocabularies, the RS has to map between them deterministically and reject the request if the mapping is ambiguous.
Section 8.5 keeps the subject key away from client authentication. If the token is sender-constrained, the RS validates DPoP or mTLS on its own terms. A good subject signature does not repair a failed client check. When the evidence names an intended presenter, the RS has to compare it against the token and the request context, and reject on mismatch.
Section 6.3 has the rule I like most in the document. If the RS introspects, the introspection response wins outright. The RS must not merge the JWK Set from the JWT with the one from introspection, and if introspection comes back without subject_keys, there is no binding, even if the token still carries one. Same kid with different key material across the two sources means reject the request and log a security event, compared by thumbprint rather than by JSON serialization. That's the kind of detail that only gets written down after someone has debugged it.
Caching in 8.6 follows from that. The RS can cache key material, but a cached key is never a current binding for a new token unless that token conveys it again.
The companion draft nobody mentions
Three of the four authors have another active OAuth draft: draft-li-oauth-delegated-authorization, now at -03. It defines Delegated Authorization Tokens, key-bound tokens that let a client delegate a constrained subset of its authority to another client without going back to the authorization server. The resource server validates an ordered token chain, the restrictions added at each step, and a DPoP proof from the leaf key.
Read the two abstracts next to each other. Both are about narrowing authority as it moves toward the resource. One does it with a chain of client-key-bound tokens. The other does it with a user-signed object verified against a key the AS publishes.
Neither draft cites the other. Not once, in either direction.
That's the question I'd want answered before either one gets adopted. If they're alternatives, the working group should pick. If they compose, and they plausibly do, since one constrains client-to-client delegation and the other constrains what the human authorized in the first place, then say so and show what the combined validation looks like. Right now two documents from the same team are solving adjacent halves of the same problem in isolation.
Significant interoperability concerns
So what actually interoperates here?
The claim itself is specified tightly enough that two implementations will agree on what a subject_keys value looks like. JWK Set, at least one key, asymmetric, public parameters only, kid required and unique, no x5u, no none, use must be sig if present. That part is testable and unambiguous.
Everything that would make it useful is delegated to an evidence profile that doesn't exist. So the interop surface of this document is "put a JWK Set in a claim spelled subject_keys." That's a registration.
The bigger gap is that nothing signals any of this.
There's no AS metadata parameter for advertising support. There's no way for a resource server to tell a client it needs a binding. Section 7.2 says a deployment MAY use scopes, resource indicators, or RAR to decide that an RS requires one, then defines none of them. So whether the claim appears is pure out-of-band configuration between an AS and an RS that already agreed offline.
That's the hole where the predecessor's RAR request used to be. Dropping the AS-signed certificate was right. Dropping the request left nothing in its place.
Error handling has the same shape. Section 10 defines no new error code and offers three ways to report a missing binding: invalid_token, insufficient_scope, or an application-specific error. Clients will see all three from different servers. This is the one condition in the document that would genuinely benefit from a code of its own, and draft-mcguinness-oauth-insufficient-claims is a better place to get it than inventing one here.
Two smaller things worth fixing:
kidSHOULD be the RFC 7638 thumbprint, or else something "collision resistant within its key-binding namespace." Make it a MUST. Selection bykidworks either way, but a mandatory thumbprint lets the RS verify for free that thekidmatches the key it's about to use.- Multi-audience tokens are permitted by 6.2, discouraged by the privacy considerations, and unhandled by 7.2, which says to determine "the intended Resource Server audience" in the singular. If
audhas three entries, which keys go in? Pick one answer.
Mechanisms that could be removed
Thirty pages for one claim and one introspection member. Most of the bulk is normative requirements on behavior no other party can observe.
- Section 7.1's eight enrollment requirements. The draft says the enrollment protocol is out of scope, then puts eight MUSTs on it. A resource server cannot tell whether the AS ran a proof-of-control ceremony with a fresh challenge. This is good advice and it belongs in the security considerations.
- Section 7.2's six ordered steps. What matters is the outcome: only active keys authorized for that audience, and the token doesn't outlive the binding. The internal ordering isn't interop-relevant.
- Section 8.6's caching rules. Local RS behavior. A sentence would do.
- Section 9.2's "MUST NOT be nested in authorization_details." Already implied by 6.1 and 6.2 defining the claim as top-level.
Cutting those loses no interoperability and makes the document reviewable in one sitting.
Writing quality
Good. The RFC 2119 keywords are used correctly and consistently, the sentences are clean, and I found no grammatical problems.
Two things it does better than most new drafts. The list in 5.1 of what the claim does not assert is exactly the right way to scope a new credential. And Section 9.1 proactively saying that one key for both proof-of-possession and subject evidence is NOT RECOMMENDED is a line most authors wait to be told to draw.
The problem is density. Ten security subsections, eight enrollment requirements, twelve profile requirements, six key-selection steps, and a great many MUSTs attached to things nobody can test. Requirements you can't observe tend to get ignored, and they make a document harder to review without making implementations more correct.
It also repeats itself. "Evidence can narrow but never widen" appears in the abstract, the introduction, 5.1, 8.4, and 11.7. Once normatively in 8.4 plus once in the introduction is enough.
This would be a better document at half the length.
Background of the authors
All four authors are at Huawei. Between them they have zero published RFCs.
- Cheng-Kang Chu has one active draft, this one, and nothing expired. New to the IETF.
- Ruochen Li has this draft, draft-li-oauth-delegated-authorization, and one expired ACME draft.
- Tieyan Li has the same two active drafts plus an expired RATS draft.
- Haiguang Wang is the exception. No RFCs either, but four expired drafts across ACME, EMU, and TLS, and a seat on the 2026/2027 Nominating Committee. That's a real participation record.
So a team that knows the IETF reasonably well and is new to OAuth specifically.
They deserve credit for the rewrite. The predecessor asked the AS to sign a user certificate and requested it through a RAR authorization details type. Both are gone. Cutting your own mechanism down to a single claim between revisions is the right instinct, and it happens less often than it should.