Review · October 13, 2026
Attestation-Based Client Authentication -12, What Changed and Is It Ready for Last Call?
The short version
Attestation-Based Client Authentication (ABCA) lets a client backend vouch for an individual client instance, like a specific install of a mobile app or wallet, without the backend ever learning which authorization server the instance is talking to.
The backend (the Client Attester) issues a Client Attestation JWT bound to a key held by the client instance. The instance presents that JWT to the AS or RS in the OAuth-Client-Attestation header, along with a fresh proof of possession signed with the bound key. That proof is either a dedicated OAuth-Client-Attestation-PoP JWT or, in "combined mode," a DPoP proof that does double duty.
It can be used as a real client authentication method (attest_jwt_client_auth or attest_jwt_client_auth_dpop) or as an extra security signal next to some other method. The driving use case is the EU Digital Identity Wallet ecosystem, where OpenID4VCI and the HAIP profile depend on it. The mechanism applies just as well to any native app that wants to prove it's a genuine instance rather than a copy with an extracted client_id.
This is squarely in charter, it's been a WG draft for years, and the core design hasn't moved in a long time. The interesting question at -12 is what changed under implementers' feet since -10, and whether it's done.
What changed since -10
The -11 and -12 revisions look like cleanup in the changelog. Several of the changes are wire-visible. If you shipped against -10, you have work to do.
Breaking changes for implementers
Header syntax moved from token68 to Structured Fields. In -10, OAuth-Client-Attestation and OAuth-Client-Attestation-PoP used the RFC 9110 token68 ABNF. In -12 both are Structured Fields whose value is an Item of type Token containing the compact JWT, and the Item MUST NOT have parameters.
The good news is that a well-formed compact JWT is a valid sf-token, so the bytes on the wire for a normal request don't change. The difference is in parsing. A strict SF parser will reject things a token68 parser accepted, and a server now has a normative basis to reject a value with a trailing ;param. If you wrote your own header parsing, check it against RFC 9651 rather than assuming it still works.
iat is now REQUIRED in the Client Attestation JWT. It was OPTIONAL through -11. Attesters that omit it will produce attestations a -12 server rejects under Section 7.1 step 2. The PoP JWT already required iat since -07, so this only affects the attestation itself. Client Attester operators need to ship this before ASes start enforcing it.
Combined mode uses only DPoP nonces now. This came in -11 and it's the biggest behavioral change. In -10, a challenge from the challenge endpoint went into the DPoP proof's nonce claim. In -11 and -12, combined mode ignores the ABCA challenge machinery entirely. Freshness comes from the RFC 9449 nonce mechanism, the error is use_dpop_nonce rather than use_attestation_challenge, and the challenge endpoint delivers a fresh nonce in a DPoP-Nonce response header instead of in the JSON body. A -10 client in combined mode that copies attestation_challenge into its DPoP nonce will fail against a -12 server.
Fetching from the challenge endpoint is no longer mandatory. -10 said that if the AS offers a challenge endpoint, the client MUST retrieve a challenge and use it. -12 makes challenge support explicitly optional, says a client without a challenge SHOULD fetch one, and lets a client reuse a challenge across PoPs unless the server rejects the reuse. This relaxes client requirements, but servers that assumed every request carries a fresh challenge need to revisit that assumption.
There's a defined error-and-retry flow with HTTP status codes. -10 defined the error codes and nothing about status codes. Now an AS that wants a challenge it didn't get MUST return 400 with use_attestation_challenge and a fresh challenge in OAuth-Client-Attestation-Challenge. An RS MUST return 401 with the error in WWW-Authenticate. Section 7.4 adds general guidance: ASes SHOULD use 400 for these errors and MAY use 401 for invalid_client_attestation, RSes SHOULD use 401. The client SHOULD retry once with the new challenge and MUST NOT retry indefinitely. This is a real interop improvement over -10, where every implementation picked its own status code.
The AS must reject combined mode if it doesn't support it. Section 7 now says a request with an attestation and a DPoP proof but no attestation PoP MUST be rejected if the AS doesn't support attest_jwt_client_auth_dpop. Previously the dispatch rule just said to apply combined-mode validation, which could lead an AS to silently accept a request in a mode it never advertised.
Key binding and profiling
-12 adds guidance in Section 10.6 that a client instance SHOULD track which key each bound artifact, like a refresh token, is bound to. Platforms can rotate keys out from under an app, and the client SHOULD be able to detect that and know which artifacts just became unusable. That's practical advice. On iOS and Android, keystore keys can disappear on OS restore or biometric changes, and the user sees a forced re-login with no obvious cause.
Section 13 (new in -11) lets profiles deviate on exactly three things: the attestation typ, the meaning of sub and how it maps to client_id, and the refresh token binding. Each deviation has to say what replaces the check it removes. That's a tight, well-scoped extension point, and it gives the wallet ecosystem room to do instance-level identifiers without forking the spec.
Other additions
- A Client Metadata section (Section 9) defines
client_attestation_signing_alg_values_supported,client_attestation_pop_signing_alg_values_supported, andclient_attestation_pop_methods_supportedas client capability advertisements, explicitly not something the server has to consult. - Section 7.3 clarifies that comparing the DPoP key to
cnfmeans comparing JWK Thumbprints or required members, not serialized JWKs. - A short note explains that RFC 7521 isn't used because the attestation travels in headers.
- Privacy considerations now cover
iatandexpas correlation factors, recommending that attesters round or randomize them. - A new Section 12.3 says claims in the attestation are only as trustworthy as the attester making them.
Structured Typewas dropped from theOAuth-Client-Attestation-ChallengeIANA registration.
Issues worth raising before or during WGLC
None of these are design problems. They're the kind of thing a last call is for.
The Challenge header now has no syntax at all. The two JWT headers got a precise SF definition. The OAuth-Client-Attestation-Challenge header lost its Structured Type: Item registration and gained nothing in its place. Section 6.2 says only that it "contain the value of the Challenge." The challenge is opaque to the client and server-generated, so this mostly works in practice. A server that puts a comma or a quote in its challenge will still break some client's header handling. Either define it as an sf-token or sf-string like the others, or say explicitly that it must be a token68-safe string.
Rounding iat fights with checking freshness by iat. Section 11.1 says the attester SHOULD round iat, giving "round down values of iat to 24:00 UTC" as the example. Section 7.1 step 6 tells the server to judge freshness "by checking the iat or exp claims." An AS that wants attestations no older than an hour can't do that with day-rounded iat. Now that iat is REQUIRED, the spec should say which claim a server should use for freshness when the attester follows the privacy guidance, probably exp.
The Section 7 dispatch rules still contradict each other slightly. Bullet two says that when an attestation and a DPoP proof are present without a PoP header, the server MUST apply combined-mode validation. Bullet three says that in exactly that case, if the AS doesn't support combined mode, it MUST reject. Bullet two needs a "supports attest_jwt_client_auth_dpop" condition so the two don't overlap.
The RS error response doesn't say which WWW-Authenticate scheme to use. Section 6.1 and Section 7.4 put the error code in WWW-Authenticate per RFC 6750, which defines the Bearer scheme. An RS that receives an attestation with a DPoP-bound token, or as a pure security signal with no token at all, has no defined scheme to put the error under. One sentence fixes it.
"Most recently received Challenge" is ambiguous for concurrent clients. Section 6 says the client MUST include the most recently received challenge. A mobile app making parallel requests can get challenges back in a different order than it sent the requests. Since the server already decides whether a challenge is single use, this could relax to "a Challenge provided by the server that the Client has not yet seen rejected."
none in the PoP methods registry is an odd fit. client_attestation_pop_methods_supported with none means "attestation optional," and omitting the parameter also means "attestation optional." The only way to express "required" is to list methods without none, and even then the client requirement is a SHOULD. This works, but it'd be cleaner to drop none and say that the presence of the parameter means attestation is expected. Not worth blocking on.
Interoperability
The core is solid. There are two PoP modes, with clear dispatch rules between them, metadata values for both, and a registry for anything future. The verification steps in Section 7 are written as checklists an implementer can follow line by line, and -11 and -12 removed the duplicated challenge check that was in -10.
Trust establishment between the AS and the Client Attester is deliberately out of scope. Section 10.8 lists x5c, jku plus kid, and kid plus client metadata as acceptable options, with a SHOULD to pick one. That's the right line for this document, since ecosystems like the EUDI Wallet will pin it down with trust lists. For the open-world case, draft-mcguinness-oauth-client-attesters is already trying to fill the gap with a client metadata parameter that names endorsed attesters, which is a sign the base spec drew the boundary in a sensible place.
MAC-protected attestations are still allowed, with a security consideration steering people toward signatures. That's a reasonable carve-out for the case where attester and AS are the same entity.
Writing quality
-12 includes an editorial pass, and it shows. Capitalization of defined terms is consistent, "NOT just the client" became "not just to the client," section titles are in title case, and several run-on paragraphs got split. It reads like a document that's had a lot of eyes on it.
The document is longer than it needs to be in a few places. Section 10.7 spends a page on trie data structures for replay detection, which is implementation advice rather than protocol. Section 12.1 then describes the same jti and challenge options again. The RATS mapping in Section 14 is useful context but could be an appendix. None of that is blocking, though it'd make the security review easier.
Background of the authors
Tobias Looker (MATTR), Paul Bastian (Bundesdruckerei), and Christian Bormann (SPRIND). All three are co-authors of the Token Status List draft, and Bastian and Bormann are also on the HTTP Message Signatures for OAuth draft. Looker has one published RFC (RFC 9597) plus work in COSE and CFRG on BBS signatures. Bastian and Bormann come out of the German EUDI Wallet implementation work, which is exactly where this spec gets exercised hardest.
Ready for WGLC?
Yes, with a short list. The design has been stable for several revisions, the -11 and -12 changes are about precision rather than direction, and there's real deployment pressure from the wallet ecosystem.
The breaking changes since -10 are the real story for anyone who has already implemented. Structured Fields parsing, a required iat, and the switch of combined mode to pure DPoP nonces all need code changes on one side or the other. Those changes are improvements. Implementers should update now, before WGLC locks them in.
The Challenge header syntax gap and the iat rounding conflict are the two items I'd want fixed before this goes to the IESG. Both are one-paragraph edits. If the authors can turn a -13 around quickly, start WGLC on that. Otherwise start it on -12 and raise them as last call comments.
Further reading: draft-ietf-oauth-attestation-based-client-auth.