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

Update review · November 20, 2026

Transactional Access Tokens -01 Update

draft-tulshi-oauth-transactional-access-tokens -01View on Datatracker →

Status: Individual draft (not adopted), Standards Track. Revision -01 published 18 November 2026, seven weeks after -00. 32 pages, up from 25. Same sole author.

This revision replaces the transaction context model. The tctx and rctx claims are gone from the token. Transaction context is now a RAR object of a new type, transaction, requested in authorization_details and returned in the authorization_details claim.

That resolves the biggest structural complaint in the -00 review. The token request rules, the handle, the typ, and the metadata are unchanged.

What got fixed

  • tctx and authorization_details no longer overlap. The -00 review said the example tctx was a RAR object and the resource server had no guidance on which to enforce. Now there is one structure. Section 3.5 says it represents "both what the client was granted and what the AS asserts about the transaction."
  • The chaining profile is now referenced. The Introduction has a table of three hops: Txn-Tokens inside an application, the chaining profile across enterprises, Txn-ATs from the AS to the application. New Appendix B describes how AS-B in a chained flow can issue a Txn-AT, which is the composition the -00 review asked for.
  • RAR is a normative reference.
  • "AS-asserted" became "AS-granted." Section 7.6 now says a client "can propose transaction context, but the AS decides what is granted." That is a more honest description of an AS deciding on the client's own account of the task.
  • The tctx immutability question is moot. The claim is no longer in the token.

What's partly addressed

  • The typ question is now an open issue. Section 3.1 still requires txnat+jwt. Appendix C asks whether Txn-ATs should be ordinary at+jwt tokens identified by the transaction type and the txn claim. The move to RAR makes the case for at+jwt stronger, and the draft says so: the plain type "lets existing RAR deployments issue and consume Txn-ATs with fewer changes."
  • Context still has no shape. The transaction type defines the RAR common fields and one new field, requester. Everything else is "defined by the deployment or by a profile." Every AS and resource server pair still needs a bilateral agreement on what task means. The draft flags the type name and field set as provisional.
  • Starting a new transaction is still ungated. Section 4.3 still says an absent txn_handle begins a new transaction. The new Durable Consent section says an AS SHOULD bound what a consent authorizes by rate, resource servers, or elapsed time. That is the right place for the gate, but it's advice in Security Considerations. Section 4.4's limits still only bind a client that sends the handle.

What didn't change

  • transactional_access_tokens_required still contradicts the grant rules. The AS MUST issue only Txn-ATs to that client and MUST NOT issue one from the authorization code grant. It's now slightly worse. A new transaction requires a transaction object in the request, so the flag also can't apply to a request that omits one.
  • Section 13.13 of Txn-Tokens is still not mentioned. The positioning table explains where Txn-ATs sit. It doesn't explain why the parent spec's "deliberate design choice" to keep transaction context out of access tokens needed revisiting.
  • "MUST evaluate policy for every Txn-AT request" is still untestable.
  • The handle expiry is still not returned. Still an open issue.
  • The Overview steps are still misnumbered.
  • DPoP and mTLS are still informative references. Section 5 has a MUST for proof of possession, and the new Section 6.1 specifies DPoP proof verification steps.
  • None of the removal suggestions were taken. The HMAC derivation is still in Security Considerations. The transactional_access_tokens_supported flag is still there. It's now redundant, because an AS that supports Txn-ATs MUST list transaction in authorization_details_types_supported.

What's new

The transaction authorization details type. Section 3.4. A new transaction request MUST carry exactly one. The Txn-AT MUST carry exactly one, and the resource server MUST check for it. Later requests with a txn_handle can omit it, and the AS carries the context forward and may tailor it for the new audience.

Enrichment. The AS adds a requester field with what it observed about the client, like the authentication method and source IP. Clients MUST NOT send it. This is where rctx went. Fields the AS doesn't understand MUST be dropped or the request rejected, which is a good rule.

Sender-constrained Txn-ATs at the TTS. Section 6.1. The resource server forwards the client's DPoP proof to the TTS, and the TTS verifies it against cnf itself instead of trusting the resource server's word.

Durable Consent. Section 7.4 admits that one consent can outlive the user's awareness of what the agent is doing. It suggests bounding consent and showing users the transactions performed under it.

A correlation warning for task names. A task identifier that is the same in every Txn-AT of a transaction defeats pairwise txn. The AS SHOULD vary or omit it per resource server.

New concerns

Section 6.1 isn't implementable yet. The draft says it "does not yet define the Txn-Token request parameter that carries the presentation evidence." The verification list is also missing the DPoP freshness checks. There is no iat window, no jti check, and no nonce handling, so a forwarded proof can be replayed to the TTS for as long as the Txn-AT is valid.

Section 6.1 may not be needed. The TTS already trusts the resource server to supply request_details. For mTLS-bound tokens the draft falls back to the resource server's assertion, "a reported assertion, not verified evidence." If that is acceptable for mTLS, the working group will ask why DPoP needs more.

The mapping into tctx is undefined. The TTS SHOULD derive tctx from the transaction object. Does it copy the whole object? Drop type and requester? Two TTS implementations will produce different Txn-Tokens from the same Txn-AT.

"Consistent with the existing transaction, as determined by the AS" is the only rule for a transaction object sent alongside a handle. A client can't predict what will be rejected.

transaction is a generic type name. RFC 9396 recommends collision-resistant names, and payment deployments are likely to already use something like it. The draft knows, and lists a URN as an option.

The resource server SHOULD now use the object in authorization decisions. It was a MAY for tctx. A SHOULD on deployment-defined fields doesn't tell an implementer what to do.

The chaining profile reference is stale. The draft cites draft-fletcher-transaction-token-chaining-profile-02. The current document is draft-fletcher-oauth-txn-token-chaining-profile.

Writing quality

Still clean. The examples moved from an appendix to Section 1.1, ahead of the terminology, which helps a first-time reader. The Document History section is specific about what changed and why.

The page count grew by seven for a revision that removed two claims. Most of that is the relocated examples, Section 6.1, and Appendix B.

Background of the authors

No change. Atul Tulshibagwale (CrowdStrike) is still the sole author and Pieter Kasselman (Defakto Security) is still the listed contributor. The acknowledgments now thank Jeff Lombardo, Girish Konda, Wes Jackson, and Ashay Raut for feedback.

Where this leaves the draft

The RAR alignment was the right call, and it changes what the draft is. A Txn-AT is now a short-lived, single-audience RFC 9068 token with a txn claim and one RAR object of a known type. That is close to something an existing RAR deployment can issue today.

So what's left that needs a new token type? That's the typ open issue, and it's the next thing to settle. The other is the field set of the transaction type, because without it this is a container format and the interop still happens bilaterally.