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

Review · November 21, 2026

Binding Agent Dialogs to WIMSE and OAuth Credentials

draft-reddy-wimse-oauth-dialog-binding -00View on Datatracker →

The Problem

AGENTPROTO is building a dialog management protocol. Every request and response will carry a dialog context that names the dialog and the participant that started it.

Nothing protects that context today. A WIMSE proof does not cover it, so a hop can rewrite it. A Transaction Token does not record it, so a token obtained in one dialog can be presented in another.

This draft ties the three together. The receiver checks who sent the request, what it may do, and which dialog it belongs to, and that the answers agree.

How It Fits In

This is the security companion to the AGENTPROTO work, written by two of the same authors. The requirements draft hands 16 requirements to OAuth and WIMSE, and this is the first attempt at meeting some of them.

It touches three working groups:

The WIMSE charter covers binding a token to a transaction and carrying context across multiple hops, so the signature parts fit. The Transaction Token extension belongs in OAuth.

The new surface is small: one covered header, one JWT claim, one tctx member, and a handful of receiver checks. It creates no new roles. It leans on six WIMSE and OAuth drafts that are all still in progress, plus one individual draft for multi-hop.

Summary

There are three bindings.

The proof covers the dialog context. With HTTP signatures, dialog-context joins the list of headers that must be signed when present. With the WPT, a new dth claim carries a hash of the header. Either way, a change in transit breaks the sender's proof.

The Transaction Token records the dialog. A workload asks the Transaction Token Service for a token and passes the dialog identifier in request_details. The service writes it to tctx.dlg. The receiver rejects the request if the token's dialog and the header's dialog differ.

The proof covers the Transaction Token. WIMSE already does this. The draft adds nothing here.

Dialog setup takes no extra round trip. The initiator proposes an identifier and names itself in its first request. The server checks that the named initiator matches the WIT that signed the request, then creates the dialog.

So what happens at a trust boundary? The dialog stops. The agent starts a new "linked dialog" in the other domain with itself as initiator. It trades its Transaction Token for a JWT grant, then for an access token from the other domain's authorization server. Neither the dialog identifier nor the Transaction Token crosses over.

Proving who started a dialog several hops back is handed to draft-reddy-wimse-aggregate-signatures.

Interoperability Concerns

The thing being signed doesn't exist. Dialog-Context is a placeholder. The draft says its name and format will follow the agent communication protocol "once that protocol is defined." The covered component name and the input to the dth hash are both undefined until then. Nobody can implement this today.

It assumes the context is an HTTP header. A2A and MCP carry their context identifiers in the message body. If AGENTPROTO does the same, Content-Digest already covers it and most of Sections 7 and 8 has nothing to do.

The dialog check trusts the sender. The receiver compares tctx.dlg to a header the sender wrote. A workload holding a token for D1 writes dlg=D1 and passes. A stateless tool has nothing else to compare against. The check catches a token that leaks to a workload outside the dialog, and that workload already needs a valid WIT to get anywhere.

The Transaction Token Service has nothing to check. It issues the token if the request passes "local policy." It has no way to know the dialog exists or that the requester is part of it. "Dialog-bound authorization" is only as good as that unspecified policy.

Linked dialogs inside a domain collide with token immutability. Transaction Tokens say tctx stays fixed through the call chain. A linked dialog gets a new identifier. The draft doesn't say how the next hop gets a token that matches, or whether a replacement token may change dlg.

The origin property depends on a draft that can't be implemented. Verifying the initiator needs the earlier hops' signatures. The aggregate signatures draft has no usable algorithm identifier yet. The receiver step also reads "if proofs of earlier hops are present," so the check is optional. Without it, the initiator field is whatever the last hop says it is.

The two proof types give different guarantees. The WPT has no response proof and no multi-hop story. A WPT deployment gets request integrity for one hop and an unauthenticated dialog setup response. The draft admits this, and still presents both as equal options.

It disagrees with the architecture draft. draft-manyfolks-agentproto-arch says each delegation creates a new dialog with a new ADID, and the delegating agent passes along the chain of prior ADIDs and the principal. Here a delegated agent keeps using D1 for its next call. At a boundary, "only A2 knows that D2 continues D1." Both authors are on both drafts. One of them needs to change.

The cross-domain section has no protocol in it. It contains no normative requirements. It doesn't say how a Transaction Token is presented to the domain's authorization server, and the Transaction Token draft warns against using these tokens outside their call chain. B1 must trust domain A's WIT issuer, which is a big assumption left in one clause. The access token B1 receives isn't tied to D2 at all.

No registry for dlg. The draft notes there is no registry for tctx members. Two extensions picking the same name will collide.

What Could Be Removed

Most of this could be three sentences in other documents.

RFC 9421 lets a signer cover any header. "Sign the dialog context header when present" is one line in the AGENTPROTO HTTP binding, or one more bullet in the WIMSE HTTP signatures list.

dth could be an entry in the WPT's existing oth object instead of a new registered claim. The draft already borrows the oth hash rules.

Section 11 restates Identity Chaining with a dialog label attached. It could be a paragraph.

Section 13 repeats Sections 7 through 11 as a list. Appendix A compares against A2A and belongs in the architecture draft.

That leaves tctx.dlg and the receiver check as the real contribution. That part needs the most work.

Writing Quality

The writing is good. Sentences are short and direct, the figures help, and the structure follows IETF conventions.

It is honest about its limits. It says the WPT leaves responses unauthenticated and that one hop's proof says nothing about the initiator.

It repeats itself. The "who sent it, what it may do, which dialog" framing appears four times. The single-hop limitation is explained in Sections 3, 6, 9 and 14.3.

Normative language is uneven. Sections 7, 8 and 10.4 use MUST. Dialog establishment, token requests, forwarding, revocation and the main receiver procedure are written in plain present tense, so it's unclear what is required.

The references have already gone stale. The requirements draft it cites has since been adopted by AGENTPROTO.

Authors

Tirumaleswar Reddy.K (Nokia) has 39 RFCs since 2014, mostly in STUN, TURN, DOTS and PCP. He has 51 active drafts, heavy on post-quantum work, and reviews for the security and ops directorates. He co-authors the AGENTPROTO requirements and architecture drafts and three other WIMSE individual drafts. Very active writer.

Zaheduzzaman Sarker (Nokia) has 7 RFCs from 2014 to 2021, all in transport and real-time media. He chairs TSVWG and TIPTOP and is a former Transport Area Director. He has 10 active drafts, six of them on agent communication.

Both are experienced IETF authors and central to AGENTPROTO. Neither has published in OAuth, and the weakest part of the draft is the OAuth part.