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

Review · November 16, 2026

WIMSE Workload Proof Token Moves Into the Authorization Header

draft-ietf-wimse-wpt -02View on Datatracker →

The Problem

A Workload Identity Token (WIT) binds a public key to a workload identity. It must not be used as a bearer token. Something has to prove the sender holds the private key.

The Workload Proof Token (WPT) is that proof. It is a short-lived JWT, signed with the WIT's key, sent with each HTTP request.

How It Fits In

This is an adopted WIMSE working group draft, not a new submission. It was split out of the old draft-ietf-wimse-s2s-protocol and is now at -02. It has a 10 minute slot in WIMSE Session A at IETF 127.

It is one of three ways to prove possession of a WIMSE credential. The other two are mutual TLS and HTTP Message Signatures. The WIT itself is defined in WIMSE Workload Credentials.

The design is borrowed from DPoP. The HTTP Message Signatures draft calls the WPT "the DPoP Inspired Option". It defines no new roles. It adds one HTTP authentication scheme, one media type, and three JWT claims (wth, tth, oth).

So what is left before publication? The document is small. It has two normative references to unfinished drafts, the WIMSE architecture and workload credentials. It has no implementation status section. The larger problem is that -02 made a breaking change that the neighboring drafts have not caught up with.

Summary

The sender puts the WIT in the Workload-Identity-Token header and the WPT in Authorization: WPT <jwt> (Section 2).

The WPT carries:

  • aud: the target URI of the request, without query or fragment.
  • exp: a short expiry, "minutes or seconds".
  • jti: a unique ID.
  • wth: a SHA-256 hash of the WIT.
  • tth: a hash of the Transaction Token, if one is in the request.
  • oth: hashes of other tokens in the request, keyed by lowercase header name.

The recipient validates the WIT first. Then it checks the WPT signature against the key in the WIT's cnf claim, and checks the audience, expiry, and hashes.

What changed in -02

One change matters more than the rest.

  • The WPT moved from its own header into the Authorization header. The Workload-Proof-Token header is gone. There is a new WPT authentication scheme, registered with IANA (Section 4.3).
  • The ath claim is removed. -01 used it to bind an OAuth access token to the WPT. The -01 example request carried Authorization: Bearer next to the WPT. That example is gone.
  • Error handling reversed. -01 said 401 was NOT RECOMMENDED. -02 says the recipient SHOULD return 401 with WWW-Authenticate: WPT (Section 2.1).
  • Section 2.2 is rewritten. -01 described accepting bearer tokens and WITs side by side during a migration. -02 says a request with a WPT cannot carry any other authentication scheme, and "This exclusivity is intended."
  • New Section 3.1.5: implementations MUST NOT log raw WPTs by default.

Interoperability

Significant interoperability concerns

A WPT request can no longer carry an OAuth access token. Authorization holds one credential. With the WPT there, Bearer and DPoP are both shut out. Section 2.2 says tokens carrying user or authorization context are fine "provided that they are conveyed in their own header fields". Transaction Tokens have their own header. Access tokens do not. A workload calling an OAuth protected resource has to pick one, or invent a header.

The HTTP Message Signatures option does not have this limit. It lists Authorization as a component the signature can cover. The two application-layer options now differ on whether they work with OAuth at all.

The sibling drafts still describe the old design. Workload Credentials -03 Section 5.4 still says the WIT "can therefore be presented along with existing headers used for JWT bearer tokens" during a transition. WPT -02 removed that same text. The AIMS draft (Section 9.2.1) still says the WPT travels in a Workload-Proof-Token header.

The audience is loosely specified. aud SHOULD be the target URI, but may be a "deployment-specific value". The recipient accepts "an acceptable alias or normalization thereof". Two independent implementations have no shared rule for what matches.

The proof is weaker than DPoP in three ways. There is no HTTP method claim, so a WPT captured from a GET is valid for a DELETE to the same URI. The query string is excluded. There is no server nonce. Replay checking on jti is only RECOMMENDED. Section 3.3 admits the WPT "does not protect major portions of the request". That is a fair trade for simplicity, but the method is cheap to add.

oth depends on knowledge the recipient may not have. It is mandatory when "other tokens" conveying user identity are present. The draft does not say which headers count. The recipient MUST reject entries it does not understand.

Mechanisms that could be removed

oth could be deferred until there is a second token type that needs it. tth covers the one concrete case.

Writing Quality

Section 2 is clear and easy to implement from. The new Section 2.2 explains its reasoning well.

The security considerations still read like text cut from a larger document:

  • Section 2 cites "Section 5.1.4" of the credentials draft. WIT validation is Section 5.1.3 in -03. No 5.1.4 exists in -02 or -03.
  • The history says DPoP moved to the informative references. It is not in the references at all.
  • Section 3.2 says the WIT "is signed by a private key in possession of the workload". The issuer signs the WIT. The workload signs the WPT.
  • Section 3.4 refers to "the Workload Identifier present in the WPT". The WPT has no identifier claim.
  • RFC 7517, RFC 7518, and RFC 7800 are normative references that are never cited.
  • The media type registration says WPTs "integrity-protect messages". Section 3.3 says they do not.

Authors

From their Datatracker profiles:

  • Brian Campbell (Ping Identity). 14 RFCs, including RFC 9449 (DPoP), RFC 8705 (mutual TLS), and RFC 8693 (token exchange). 9 active drafts across OAuth, WIMSE, and JOSE.
  • Arndt Schwenkschuster (Defakto Security). No published RFCs. 8 active drafts, including WIMSE Workload Credentials and OAuth SPIFFE Client Authentication. He is presenting this draft at IETF 127.

Questions for the Working Group

So what should the session decide?

  • Is the working group sure about the Authorization header? This is the last cheap moment to reverse it. If it stays, say plainly that WPT and OAuth access tokens cannot be used together.
  • If they cannot, how does a workload present a WIT to an OAuth protected resource? Is the answer "use HTTP Message Signatures"?
  • Should the WPT include the HTTP method?
  • Can the audience matching rule be made exact?
  • Who updates Workload Credentials Section 5.4 and AIMS to match?