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

Review · November 21, 2026

Message Binding for Workload Proof Tokens

draft-nandakumar-wimse-wpt-message-binding -00View on Datatracker →

The Problem

The Workload Proof Token proves a workload holds the key in its Workload Identity Token. It is only fully specified for HTTP. Its aud is the request URI, it rides in the Authorization header, and it does not cover the request content.

Agent protocols are starting to run over Media over QUIC, WebSocket, and message brokers. There is no request URI to bind to. The relays in the middle terminate TLS and do not verify anything.

This draft defines how a WPT travels with a message on any transport. It adds a hash of the message content to the proof, plus a verifier-supplied nonce for freshness.

How It Fits In

This fills a gap the WPT draft names itself. Its introduction leaves "detailed bindings for non-HTTP protocols" to future specifications. This is the first attempt at one, so it sits squarely in WIMSE's scope.

It is a small extension. The normative dependencies are the WPT and Workload Credentials drafts, plus JCS (RFC 8785) and SD-JWT (RFC 9901). It adds one new JWT claim, reuses the registered nonce claim, and creates one IANA registry. It defines no new roles.

It generalizes the dth claim from draft-reddy-wimse-oauth-dialog-binding, which hashes one HTTP header. Here the hash covers a whole message.

It also reaches back into HTTP. A WPT on an ordinary HTTP request may carry the new claim over the request body. That overlaps with the WIMSE HTTP Message Signatures profile, which already covers content through Content-Digest.

How It Works

The application protocol defines a "credential member" in its messages. That is where the WIT, the WPT, and any other tokens go. The draft does not define it.

The WPT changes in three ways:

  • aud is the receiver's workload identifier, or any other value the application protocol picks, instead of a URI.
  • mth is required. It is an object with a canonicalization profile name p and a SHA-256 hash h of the message under that profile.
  • nonce echoes a value the verifier handed out earlier. The verifier accepts each nonce once.

The existing wth, tth, and oth claims carry over, with "request" read as "message".

One profile is defined, json-jcs. Remove the credential member from the JSON object, serialize the rest with JCS, and hash it. A registry lets other documents add profiles for other formats.

The receiver recomputes the hash and rejects on mismatch or on an unknown profile. A relay can forward and even re-serialize the JSON, but it cannot change the content or move the proof to a different message.

Significant Interoperability Concerns

Nothing here interoperates on its own. The credential member, the aud value, the oth key names, and nonce delivery are all left to the application protocol. That is a reasonable split for a framework. But no application protocol is named, so there is no way to test whether the framework is sufficient.

json-jcs is not a complete profile name. The octets depend on which member gets removed, and p does not say. The verifier has to know out of band. The draft then offers a second option: a protocol MAY register its own profile that is "json-jcs with that member". Two ways to name the same computation is one too many. Pick one, or add the member name to mth.

The credential member is a hole in the hash. Everything in it is unsigned unless separately listed in wth, tth, or oth. The draft should say that a verifier ignores anything else found there. It also doesn't say whether the member must be at the top level of the object.

The motivating case has no profile. The draft opens with Media over QUIC relays. Objects on a media track are binary, and the only profile is for JSON. Section 5 describes what a binary profile would do but does not define one. A plain "hash the octets as delivered" profile would cost three sentences.

The nonce assumes a conversation. A verifier must have sent something to the signer first. That works for request and response. It does not work for the first message in an exchange, or for publish and subscribe with many receivers, which is the broker and relay case the draft is aimed at. The requirement levels also don't match: the signer SHOULD include the nonce, the verifier MUST reject without it.

Replay rules are stricter than the base spec. The WPT draft makes the jti check RECOMMENDED and admits that syncing a replay cache across instances is hard. This draft says a multi-instance verifier MUST share its record of nonces and jti values and MUST reject when it can't reach it. That is a new availability dependency. It is unclear whether it applies to HTTP WPTs that carry mth.

JCS in front of authentication. On HTTP, the verifier has to parse and canonicalize the whole body before it knows the proof is valid. JCS only works for I-JSON, so integers beyond 2^53 break it. The draft requires rejecting duplicate member names, and many JSON parsers silently drop them. The named targets are MCP and A2A, where payloads are large and arbitrary.

An HTTP verifier that doesn't know mth ignores it. The sender believes the body is bound. The receiver never checks. There is no way for a sender to learn which it is talking to.

Mechanisms That Could Be Removed

  • The Idempotency-Key rule in Section 5.2. It puts a MUST on requests signed with HTTP Message Signatures, in a document about the WPT. It belongs in the HTTP signature profile if anywhere.
  • The SD-JWT section. It is three sentences that let a Key Binding JWT carry mth. SD-JWT is OAuth working group territory, and the section says nothing about how verification works there.
  • The clock and Roughtime paragraph. Every JWT with exp has this problem. It is not specific to message binding.
  • The idempotency key guidance. Repeated operations are an application concern, as the draft itself says.
  • mth on HTTP requests, or at least split it out. It is a different feature from non-HTTP binding and competes with adopted work. It deserves its own discussion.

Writing Quality

The writing is good. Sentences are short and precise, IETF conventions are correct, and the body is about five pages. It states what each mechanism does not do, which most drafts skip.

It is too short in places. There is no worked example of a message, its canonical form, and the resulting WPT. For a canonicalization spec, that is the most useful thing it could include. aud and the credential member each get one sentence.

The related work section is a fifth of the body. It positions the draft well against DPoP, CAT-4-MOQT, and dialog binding, but some of that space would be better spent on the example.

Authors

Suhas Nandakumar (Cisco) has 9 published RFCs from 2013 to 2025, all in real-time media: STUN and TURN URI schemes, RTP taxonomy, SDP and ICE. He has 36 active Internet-Drafts. Most are Media over QUIC, where he is an editor of the transport draft. About ten are recent OAuth, WIMSE, and AGENTPROTO drafts. He is an experienced IETF author who is new to this area and writing at high volume right now.

Cullen Fluffy Jennings (Cisco) has 29 published RFCs from 2002 to 2026, including SIP identity (RFC 3325, RFC 4474) and NAT behavior (RFC 4787). He has 30 active Internet-Drafts, mostly Media over QUIC plus the same set of agent, OAuth, and WIMSE drafts co-authored with Nandakumar. He is a long-time IETF contributor, also new to these working groups.

The Media over QUIC background explains the framing. This is one of a set from the same pair, alongside their delegator-bound chains draft.

Bottom Line

This is a real gap, and the WPT draft says so itself. A content hash under a named profile is the right shape for the fix, and the draft is small enough to review in one sitting.

So is it ready? Not yet. The one profile it ships can't be computed from the claim alone, and the transport it was written for has no profile. The nonce needs a story for one-way and one-to-many messaging.

The useful next step is a concrete application protocol that uses it, with a worked example. The HTTP body binding should be argued separately, against Content-Digest in the WIMSE HTTP Message Signatures profile.