Review · November 21, 2026
Test Vectors for the WIMSE HTTP Message Signatures Protocol
The problem
The WIMSE HTTP Signatures profile has a long list of MUSTs and only a few worked examples. Two implementations can both claim conformance and still disagree on details like the value of @query when there is no query string.
This draft publishes a fixed set of messages, keys, signature bases and expected outcomes so implementers can check their code against the same inputs. It defines no protocol behavior.
How it fits in
This is a companion to draft-ietf-wimse-http-signature, the WIMSE profile of HTTP Message Signatures (RFC 9421). It is in charter in the sense that it supports an adopted WIMSE document. It adds no mechanisms, roles or extensions.
It depends on one working group draft and inherits that draft's dependency on WIMSE Workload Credentials for the Workload Identity Token (WIT).
The draft pins the base spec at -08. The base spec is already at -09, and that matters:
- Observation O-2 says the base spec's example nonces are below the entropy minimum and that the entry stays "until the note appears in the text." The note is in -09.
- -09 also clarified the CONNECT
@authoritydefault-port rule that P-04 was built to test. - -09 lets a client reuse a server-generated nonce from
Accept-Signature. None of the vectors cover that.
The set started as a comment on a base spec issue on 6 November and went through four revisions in a week before becoming this draft. That history is the best thing about it. The feedback loop already changed the base spec.
What's in it
Three services (svcA, svcB, svcC) and a WIT issuer, with all private keys published. Keys are derived from fixed seeds. Signatures use Ed25519 and deterministic ES256, so every value is reproducible.
Six positive vectors:
- A signed GET that asks for a signed response, and the signed response
- A POST with a body, signed with ES256
- A CONNECT request, which covers
@authorityinstead of@pathand@query - A request with no query string
- A message with two signatures, where the verifier must pick the WIMSE one by
tag
Twenty-three negative vectors, each quoting the base spec requirement it tests. Most are a single mutation of a positive vector: drop expires, add keyid, add a trailing slash to wimse-aud, send a mismatched wimse-req-nonce.
Three observations, flagged for the working group and not scored.
Significant interoperability concerns
For a test vector document, interoperability means one thing. Can a stranger reproduce the bytes from the document alone? Right now the answer is no. This review did not recompute the signatures, and the problems below are visible from the text alone.
It is unclear which bytes were signed. Every signature base shows the WIT line as a placeholder. The conventions section says the placeholder is "retained verbatim in the signed material" and also that "deterministic expansion substitutes the full token text." Those two statements point to different signature inputs. An implementer has to guess, then try both.
The placeholder in the artwork is not the one the text defines. The text describes [WIT-x, full text in the Test Key Material section]. The vectors contain [WIT-A,§2 全文], untranslated, and pointing at Section 2 when the tokens are in Section 3. If the placeholder really is signed verbatim, the exact bytes matter, including the full-width comma.
One fold breaks a vector. The draft uses the RFC 8792 single-backslash fold, which strips leading whitespace on the continuation line. In P-06 a fold lands directly before a significant space in Signature-Input. Unfolding by the stated rule yields "@query""workload-identity-token" with no separator.
Nine negative vectors have no signature value. The reader is told to apply the mutation and re-sign. That makes them recipes, and N-21 has no base message at all, only "the P-02 form migrated to a CONNECT exchange." N-15 promises a literal message and signature base and prints only two headers.
Several negatives can't isolate the check they name. N-01 and N-05 change Signature-Input and keep the old signature. A verifier that never checks coverage or required parameters still rejects them, because the signature no longer verifies. The draft handles this correctly in N-06 by re-signing. The same treatment is needed everywhere.
Some vectors quietly add requirements. N-09b requires acceptance at exactly expires and rejection one second later. The base spec says only that signatures outside the window are rejected. It does not say the interval is closed, and it says nothing against clock skew tolerance. N-01 and N-08 state that signature verification "is not reached," which is an ordering the base spec does not impose. These may be the right readings. They belong in the base spec as text, not in a companion document as expected results.
Mechanisms that could be removed
Six entries are not vectors: N-16 through N-20, and N-22. They describe a misbehaving implementation and ask for a "conformance judgment."
N-16 asks a harness to estimate nonce entropy from a sample of requests. N-20 falls back on "implementation inspection or instrumentation evidence." N-22 is the failure case of P-06 restated. That is a conformance test plan, and it should either be split out or dropped.
The observations section reads like issue tracker content. O-2 is already resolved. O-1 and O-3 are fair points for the mailing list.
The provenance material can go too: the Zenodo DOIs, the v1 through v4 change log, who confirmed what on which date. None of it helps an implementer run a vector.
So does this need to be its own RFC? Probably not. RFC 9421 carries its test vectors in an appendix. The six positive vectors would fit in the base spec the same way, and the negatives fit better in a machine-readable file in the working group's repository. The draft mentions a machine-readable layer in the Zenodo dataset but includes none of it. A separate document pinned to -08 will keep drifting from the base spec.
Writing quality
The prose is clear and the structure is consistent. Each negative vector has the same three parts: construction, requirement, expected outcome. Quoting the exact MUST being tested is a good habit.
The draft is less polished in the details. The introduction points to the change log in Appendix A, and it is Appendix B. The abstract says each of the twenty-three negative cases maps to a normative requirement, though several map to behavior a verifier can't observe. The count of "23 negative vectors" includes N-09b, which is half positive.
Length is reasonable at 23 pages. It would be shorter and more useful with the behavior cases removed and real bytes in place of the re-signing recipes.
Author
Jiahui Miao (Vertciti) has no RFCs. This is their only Internet-Draft, active or expired. They are a new IETF author.
The acknowledgments credit a base spec co-author for review and three outside implementers for reproducing earlier versions of the set.
Bottom line
The work behind this draft is useful and has already improved the base spec. The document is not yet a reliable reference. Fix the placeholder ambiguity, publish literal bytes for every vector, and re-sign the negatives so each one fails for a single reason.
The better home for the positive vectors is an appendix of draft-ietf-wimse-http-signature, where they stay in step with the text they test.