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

Review · October 13, 2026

Taking a Transaction Token Across the Trust Boundary

draft-fletcher-oauth-txn-token-chaining-profile -00View on Datatracker →

The problem it's solving

Transaction Tokens carry the authorization context of a request through a company's internal call chain. They are deliberately scoped to one trust domain. The Txn-Token spec says a Txn-Token MUST NOT be accepted outside the trust domain named in its aud.

So what happens when a workload deep in that call chain needs to call a partner API?

That's the question this draft answers. The running example is good: an enterprise mail server receives a message over SMTP, mints a Txn-Token for the delivery, and the storage service needs to call a third-party spam rating API before writing the message to the mailbox. The storage service can't send its Txn-Token to the spam vendor. It needs a credential the spam vendor's authorization server will accept, carrying just enough context to be useful.

The answer is Identity and Authorization Chaining with the Txn-Token as the subject token. The workload exchanges its Txn-Token at its own AS for a JWT authorization grant, then presents that grant to the partner's AS for an access token. The Txn-Token never leaves the building.

The interesting part is the input. An ID token only exists when a human logged in. A Txn-Token exists for every transaction, including ones started by a mail server or a cron job. That makes this a chaining profile that works for system-initiated and workload-initiated requests, not just users.

How it fits in the working group

This is squarely in charter. It's a profile that glues together two adopted drafts, both of which share authors with this one. Pieter Kasselman is on the chaining draft and Txn-Tokens, and George Fletcher is on Txn-Tokens.

The chaining draft intentionally leaves the subject token type open and expects profiles like this one. ID-JAG is the first such profile, and this draft positions itself as the sibling. ID-JAG covers the case where the target AS already trusts a shared IdP for SSO. This draft covers the case where the two ASs have a bilateral "Cross-Domain Trust Agreement" and the principal might not be a person at all.

No new grant types, no new endpoints. It depends on Token Exchange, the JWT grant profile, RFC 8707, RFC 8414, and the two drafts. New registrations are small: a txn-chain+jwt typ value and a txn_claims JWT claim.

What the mechanism actually does

Four parties matter: the Requesting Workload in domain A, AS-A, AS-B, and the partner's resource server.

The workload sends a token exchange request to AS-A with subject_token set to its Txn-Token, subject_token_type of urn:ietf:params:oauth:token-type:txn_token, audience set to AS-B's issuer, and optionally resource set to the partner API. It authenticates as itself. The draft borrows ID-JAG's convention for audience versus resource, which is a nice bit of consistency.

AS-A validates the Txn-Token, maps the sub into an identifier agreed with domain B, and mints a JWT grant with typ: txn-chain+jwt. The grant carries iss, sub, aud, a short exp (60 seconds preferred), jti, scope, the original txn identifier, and an optional txn_claims object holding a minimized subset of the Txn-Token's scope and rctx.

The workload presents that grant to AS-B using RFC 7523. AS-B checks the typ, the issuer against its trust agreement, single-use jti, resolves the subject, optionally evaluates txn_claims, and issues an access token with no refresh token.

Section 3.3 notes that this can repeat. Domain B's resource server can mint its own Txn-Token from the inbound access token, and if a workload in B calls domain C, B plays the role of A. Trust stays strictly pairwise.

The scope rule breaks the draft's own example

Section 7.2 says the grant's scope MUST be the intersection of the Txn-Token's scope and the requested scope. Section 4.3.2 and Section 9.5 repeat that AS-A must never exceed the Txn-Token's scope.

Now look at the example. The Txn-Token's scope is mail-delivery. The requested scope is spam.rating.read. The grant contains spam.rating.read. The intersection of those two sets is empty.

This isn't a typo. It's a category problem. Txn-Token scope is the purpose of the transaction in domain A's vocabulary. The scope in the grant is a permission at domain B's resource server, in domain B's vocabulary. Those two namespaces will essentially never overlap. Every one of the three use cases in Appendix A has the same mismatch: watchlist-update becomes market data access, telemetry-aggregation becomes analytics API access.

What the draft actually needs is a mapping rule. Something like "AS-A MUST only issue grant scopes that the Cross-Domain Trust Agreement permits for the Txn-Token's purpose." That keeps the anti-escalation property without requiring a set intersection that can't succeed.

Who is AS-B actually issuing a token to?

The grant's sub is the Initiating Principal. In the user case that's the end user. In the mail case it's the mail gateway's system identity.

The workload that's actually making the call isn't in the grant. There's no client_id. There's no act claim. The draft explicitly says actor_token is not used. Appendix B frames dropping ID-JAG's required client_id as a feature, since the workload doesn't need to be registered at AS-B.

The result is that AS-B issues an access token for "this user, via domain A" without knowing which of domain A's hundreds of services it's handing it to. For the spam vendor that might be fine. For a partner applying per-client policy or investigating abuse, it's a real gap. The cnf claim helps if sender constraining is used, but that's a SHOULD and a key isn't an identity.

Txn-Tokens solve exactly this problem inside the trust domain by keeping the requesting workload identifiable. It would be a shame to lose that at the boundary. A required act claim carrying a domain-B-meaningful identifier for the requesting workload seems like the right fix.

Transaction identifiers and correlation

Section 7.1 says AS-A MUST copy txn from the Txn-Token into the grant. Section 3.3 says deployments wanting to limit linkability SHOULD generate independent transaction identifiers per trust domain. Section 10 says operators SHOULD weigh whether correlation is worth the privacy cost.

Those can't all be followed. If the MUST stands, there's no room to generate independent identifiers. Pick one.

This is where the new Transactional Access Tokens draft is relevant. Pieter Kasselman is a contributor on that one too, and it picks the opposite answer: the AS derives a different, uncorrelatable txn value for each audience from an internal transaction identifier that never leaves the AS.

The two drafts are mirror images. This draft takes a Txn-Token out of a trust domain. Txn-ATs bring transaction context into one, as an access token audienced to a single resource server that it can then exchange for a Txn-Token in its own domain. Put them together and you get the recursive flow from Section 3.3: AS-B could issue a Txn-AT instead of a plain access token, and the partner's TTS takes it from there.

Neither draft mentions the other. They should, and they should agree on how txn crosses a boundary.

Significant interoperability concerns

The most important question for a profile is which interface it makes interoperable. Here, the token exchange between the Requesting Workload and AS-A happens entirely inside domain A. Often the TTS and AS-A are the same product, and the draft allows co-locating them. That leg doesn't need much standardizing, since one organization controls both ends.

The leg that crosses organizations is the grant presented to AS-B. That's where interop matters. And the parts AS-B consumes are mostly deferred to the Cross-Domain Trust Agreement:

  • how sub is mapped, for every principal type
  • which claims may appear in txn_claims and what they mean
  • whether txn is propagated further
  • how AS-B learns AS-A's keys and issuer
  • whether just-in-time provisioning is allowed

txn_claims is the core of what makes this profile different from a generic chaining grant, and its contents are "a curated subset of Txn-Token claims" with semantics set bilaterally. Two partners will interoperate because they negotiated it, not because they implemented the spec. Some of that is unavoidable for bilateral federation. But a few defined members for common cases, like the transaction purpose, would give AS-B something it can act on without reading a contract.

It's also fair to ask whether AS-B should care that the input was a Txn-Token at all. The typ of txn-chain+jwt tells AS-B what AS-A consumed internally. AS-B mostly needs to know who the subject is, who's calling, and what's allowed. That's close to what an ID-JAG already says. The working group should discuss whether one grant format with an optional transaction context would be better than a new typ for every possible subject token.

A few concrete bugs:

  • Section 8 says an AS MUST list urn:ietf:params:oauth:token-type:txn_token in identity_chaining_requested_token_types_supported. That metadata advertises requested token types. The Txn-Token is the subject token type here, and the requested type is jwt. There's no chaining metadata for supported subject token types, so either the draft needs to define one or drop this section.
  • Section 4.1 says the Txn-Token's aud MUST identify AS-A. Txn-Tokens are audienced to the trust domain, not to a specific service. The parenthetical escape hatch covers it, but the primary rule contradicts the base spec.
  • Normative references are stale: chaining -11 (current is -17), Txn-Tokens -08 (current is -11), ID-JAG -03 (current is -04). The section numbers cited into those drafts should be rechecked.
  • DPoP is recommended in Section 9.2 with no reference to RFC 9449.

Mechanisms that could be removed

Section 3.1 and the long Section 4.2 mostly restate how Txn-Tokens work. A sentence and a reference would do.

Section 9.3 asks the Requesting Workload to authenticate to AS-A and keep the Txn-Token inside domain A. Both are already requirements in the base drafts.

The requested_token_type handling in 4.3.2 is fine, but the rule that AS-A MUST still produce a profiled grant when it's absent is redundant with the chaining draft's own behavior.

Writing quality

Good. The document reads clearly, the diagram is useful, and the three use cases in Appendix A make the case for non-human principals better than the introduction does. BCP 14 language is used properly. The AS-A and AS-B processing rules as numbered lists are easy to implement against.

Some leftover rough edges from the rename. This is -00, but the document history has entries for "Since Draft 02," "Since Draft 01," and "Since Draft 00," inherited from an earlier name. The acknowledgements still say "The author" singular, despite three authors. The draft is about the right length for a profile, and would be shorter still after trimming the restated Txn-Token background.

Background of the authors

George Fletcher shows no RFCs on his Datatracker profile, three active drafts (this one, Transaction Tokens, and First-Party Apps), and five expired drafts. The RFC count undersells him. He's a long-time OAuth and OpenID participant and one of the originators of Transaction Tokens.

Pieter Kasselman has one RFC, the recently published RFC 10027 on cross-device flow security, and seven active drafts, including Identity Chaining, Transaction Tokens, SPIFFE client authentication, and WIMSE AIMS. He's one of the most active authors in this space right now, and he's on both of the documents this one profiles.

Sean O'Dell has no RFCs and no other drafts. This is his first IETF document. Having someone from CVS Health on it is a good sign that there's a real deployment behind the use cases.

This is a credible team profiling their own specs, and the core idea is sound. The scope intersection rule and the missing requesting workload identity are the two things to raise on the list first. After that, the conversation worth having is with the Transactional Access Tokens draft, since these two documents together describe transaction context crossing a boundary in both directions and currently disagree about how.