Review · August 31, 2026
OAuth Profile for Delegated AI Agent Authorization
The problem it's solving
AI agents call protected APIs for human users. The author's position is that this doesn't need new OAuth mechanisms, it needs fewer choices.
That's the unique framing here. Most of the agent drafts in the OAuth WG right now are adding something: a new claim, a new delegation format, a new token type. This one subtracts. It takes existing OAuth and marks the optional parts mandatory, on the theory that agent instances spawn dynamically, reach high-impact APIs, and hand work off to other instances, so the loose parts of OAuth are where deployments will get hurt.
The document calls itself "DAAP" but is explicit that this is just a label, not a new framework.
How it fits in the working group
This is a security profile, not an extension. Fifteen normative references, all existing RFCs. Zero IANA actions. No new claims, no new endpoints, no new roles.
The roles map straight onto RFC 6749. An "Agent Client Instance" is an OAuth client. A "Principal" is the resource owner. The draft is careful to say the client_id identifies a registration, not a human or legal identity.
The closest existing work is RFC 9700, the OAuth Security BCP, which this draft profiles rather than extends. Functionally it's very close to what FAPI 2.0 does for financial APIs: mandate PAR, mandate PKCE S256, mandate sender-constrained tokens, mandate exact redirect matching. The draft doesn't mention FAPI, which is a gap the WG should raise.
Section 12 does the right thing on overlap. It points cross-domain delegation at identity chaining instead of defining a competing chain format, and says transaction tokens may complement rather than compete. That kind of positioning is rare in this cluster of drafts and it's welcome.
Compared to -01 this is a different document. The -01 revision was 50 pages and included a custom scope registry, budget controls with a bdg JWT claim, SSE and WebSocket event streaming endpoints, a credential vault API, a hash-chain audit log format, and DID-based agent identity with a new agt claim. All of that is gone. -02 is 19 pages and moves those topics to an explicitly out-of-scope list. The author revised hard in response to feedback and the result is much more reviewable.
What the profile actually requires
Registration binds four things: exact redirect URIs, the resources the client may request, the scopes the client may request, and key metadata for sender constraint.
Authorization is tightened on every axis at once. PAR is required. PKCE S256 is required. State is required. Exact redirect URI matching, with prefix, wildcard, and substring matching explicitly banned. Exactly one resource indicator.
Consent is where the draft spends its strongest language. The authorization server must authenticate the Principal independently of the developer or the agent. A developer API key is not consent. A policy engine decision is not consent. An automated policy may deny or narrow a request but may not approve it on the user's behalf. The security considerations put the reason plainly: conflating these lets the party requesting authority approve its own request.
Access tokens are RFC 9068 JWTs with a single audience and a cnf claim from either DPoP or mTLS. The authorization server must not emit cnf just because someone uploaded a public key, it has to prove possession first. The resource server has a seven-item validation checklist and is told that signature verification alone is not authorization.
Delegation is RFC 8693 Token Exchange with the standard act claim. Delegated tokens can only attenuate: no later expiry, no broader scope, no broader resource, and the Principal stays in sub. Depth limits and cascade revocation are SHOULDs, and the draft says parent identifiers and cascade algorithms are implementation details, not claims it defines.
Refresh tokens must be both rotated and sender-constrained. Then there's Section 7.1, which is the one genuinely new mechanism in the document.
The lost-response recovery mechanism
Refresh token rotation has a real failure mode. The server commits the rotation, then the network drops the response. The client now holds a refresh token the server considers spent, and a retry looks like token theft.
Section 7.1 defines a bounded fix. The client generates a high-entropy recovery key before its first attempt and reuses it for every transport retry of that same logical refresh. The server binds recovery state to the client, the old refresh token, and the recovery key, then returns the exact committed tokens. It must not mint anything new, extend a lifetime, or rotate again. The window is capped at a recommended 300 seconds. A wrong key, an expired window, or an already-used child token is treated as reuse and fails.
This is the most carefully specified part of the draft and it has nothing to do with AI agents. Refresh rotation lost-response recovery is a generic OAuth problem that every rotating deployment has. If the WG wants this, it belongs in its own short document that any client can use.
Significant interoperability concerns
The draft says it defines no new claims or endpoints and requests no IANA actions. Taken together with the rest of the document, that means there is nothing here for a client and server to negotiate. Two implementations cannot signal to each other that they conform to this profile. There's no metadata parameter, no scope value, no cnf variant, nothing discoverable. A client walking up to an unknown authorization server has no way to find out whether the profile is in effect, and a resource server has no way to know the token in front of it came from a profile-conformant flow.
For a document whose entire value is "these options are now mandatory," that's the central problem. FAPI handles it with out-of-band certification and a conformance suite. This draft has neither, and doesn't say what it expects instead.
The recovery mechanism in Section 7.1 is worse off. The draft specifies seven safety properties and then says the transport syntax for the recovery key "is an implementation extension and is not part of core conformance." Appendix A reveals the reference implementation puts it in an Idempotency-Key header. So the one new mechanism in the document is, as written, definitionally non-interoperable. Any two implementations will pick different wire formats. Either specify the parameter or drop the section.
"Exactly one resource value" is stricter than RFC 8707 allows. That's a defensible FAPI-style choice, but it breaks clients that legitimately need one token for multiple audiences, and the draft doesn't discuss the tradeoff.
The bigger structural question: strip out the parts that apply to any high-risk OAuth client and almost nothing agent-specific remains. There's no agent instance attestation, no way for a resource server to learn it's serving an agent rather than a browser, no signal about delegation depth, no agent identity of any kind. The draft names "Agent Client Instance" as a term and then does nothing with it that a plain client_id doesn't already do. That's arguably the honest answer to the agent authorization question, and if so the draft should say that out loud, because right now it reads as though the agent-specific content is coming and it never arrives.
Mechanisms that could be removed
Section 3.1, Optional DID Association, is optional, carries no interoperability weight, and the draft itself says DID support isn't required. It's the last remnant of -01's DID identity model. Cut it.
Section 8, Operational Extensions, is a list of things the document doesn't specify, plus advice about using JCS and RAR if you happen to build those things. Useful context for reviewers, not normative content. The requirement that "an implementation MUST NOT claim that an operator-signed hash chain is operator-independent evidence" is a normative rule about marketing claims, which doesn't belong in a protocol document.
Appendix A maps profile operations to the Grantex reference implementation's proprietary paths. Appendix B lists four possible future documents. Both are repository material.
Writing quality
Clean IETF language. Correct BCP 14 usage. Requirements are stated as testable conditions rather than aspirations. The security considerations are organized by attack rather than by section, which makes them actually reviewable.
The document is appropriately concise now, which is a large change from -01. The changelog in Section 1.3 is specific and honest. Section 13 goes further and lists exactly where the author's own implementation fails to conform: custom endpoints instead of standard ones, no PAR, DPoP proofs not validated at resources, legacy registrations that skip profile metadata. Admitting your reference implementation isn't conformant is unusual and it builds trust.
Two mechanical problems. The author's address appears twice, once as Appendix C and once as the standard Author's Address section. And the draft filename says "mishra" while the byline says "S. Kumar," which will confuse anyone tracking it in the datatracker.
The document is labeled Informational while containing dozens of MUST requirements about protocol behavior. Early drafts get this wrong routinely, so it's not worth much, but a profile like this would normally target BCP.
About the author
Sanjeev Kumar, Orchestrum Technologies LLP. Sole author, no co-authors.
Zero published RFCs. One active Internet-Draft, which is this one. No expired drafts. This is a new IETF participant with no prior document history.
The -01 to -02 revision is the thing to weigh here. Going from a 50-page bespoke protocol with credential vaults and WebSocket endpoints to a 19-page profile of existing RFCs, and reporting your own implementation as non-conformant while doing it, is exactly the response the WG hopes for and rarely gets. The author is clearly reading feedback and acting on it.
Where the working group should push
The profile-not-extension framing is the right instinct and worth encouraging, whatever happens to this specific document. The WG could use a clear statement that agent deployments should be running hardened OAuth before reaching for new claim vocabularies.
Three questions to put to the author:
- How does a client or resource server discover that this profile is in effect? Without an answer, the profile can't produce interoperable implementations no matter how tight the requirements are.
- What is agent-specific about this profile, and should the answer be "nothing, and that's the point"?
- Should Section 7.1 become its own draft? Lost-response refresh recovery is a real gap in OAuth today and it has no connection to agents.
There's also an overlap conversation to have with RFC 9700 and FAPI 2.0. Much of what this draft mandates is already covered there, and the WG should decide whether the delta justifies a separate document or a short applicability statement pointing agent deployments at existing profiles.