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

Review · October 13, 2026

Client Attester Endorsement, the Missing Link Between ABCA and CIMD

draft-mcguinness-oauth-client-attesters -00View on Datatracker →

This is one of two companion drafts published together on top of Attestation-Based Client Authentication (ABCA). The other is Client Instance Identification, reviewed in a separate post today. This one answers "who may attest for this client?" The other answers "which instance is this?"

The problem it's solving

ABCA lets a Client Attester vouch for a Client Instance and the key it holds. Before the AS relies on that, it needs to know two things: is this attester trustworthy, and is this attester allowed to speak for this client?

ABCA punts on both. Section 10.8 says trust establishment and key resolution are out of scope, then lists a few options. One of them is "kid combined with Client Metadata," pointing at the client's own jwks_uri.

That works when the AS has a bilateral relationship with every client and configures each client-to-attester mapping by hand. It falls apart for clients identified by a Client ID Metadata Document. A CIMD client shows up at an AS it has never talked to. There's no admin console where someone typed in "this client's attester is X." The AS has a URL and a JSON document, and nothing in that document says which attesters are allowed to vouch for the client's instances.

This draft adds one client metadata parameter, client_attesters, to fill that slot.

How it fits in the working group

This is squarely in charter. It extends ABCA, which is a WG document at -12, and it builds on CIMD, also a WG document. It adds no new credential, no new client authentication method, and no new grant type.

Registrations are small:

  • client_attesters client metadata parameter
  • client_attester_endorsement_supported AS metadata boolean

Normative dependencies are ABCA, CIMD, and the usual JOSE and OAuth RFCs. Both drafts reference ABCA -11. ABCA -12 changed iat in the Client Attestation from OPTIONAL to REQUIRED, and the Appendix A example here omits iat, so it's no longer a valid attestation. That's an easy fix in -01.

The draft is the successor to the instance_issuers metadata in the author's earlier Client Instance Assertion draft. Splitting the "who may attest" question out from the instance identity question, and riding on ABCA instead of inventing a parallel assertion, is a good change in direction.

What the mechanism actually does

The client publishes a list of endorsed attesters in its metadata:

"client_attesters": [
  {
    "issuer": "https://attester.example/tenant/acme",
    "jwks_uri": "https://attester.example/tenant/acme/jwks"
  }
]

The attester puts iss, sub (the client_id), and kid in every Client Attestation. ABCA doesn't require iss. This profile does, because the AS needs the issuer to pick the right entry before it can resolve kid.

The AS accepts an attestation only when both of these hold:

  1. The client's metadata currently endorses the attestation's issuer.
  2. AS policy permits that attester for that client.

The draft frames this as two separate authorities. The client publisher decides which attesters may speak for it. The AS decides which of those it's willing to trust. Neither one can act alone, and the AS can never add an attester the client didn't endorse.

For key trust, there are two policies:

  • AS-configured attester trust. The AS already trusts this attester and has its own configured key source. The endorsed jwks_uri must string-match that configured source, but the AS never fetches it. The endorsement is a permission slip, not a trust anchor.
  • Publisher-authorized key selection. The AS trusts a specific publisher (by exact client URL or HTTPS origin and path prefix) to pick both the attester and its keys. The AS fetches the endorsed jwks_uri, which must be same-origin with the issuer.

Withdrawal is just removing the entry from metadata. It takes effect once the AS's cached copy expires, and it doesn't revoke existing grants.

Does this fill a real gap?

Yes. It's the most obvious missing piece in deploying ABCA with CIMD.

Think about a desktop or mobile app that identifies itself with a CIMD URL, which is exactly what MCP clients are doing today. It can't hold a shared secret. With private_key_jwt and a jwks_uri in the CIMD, every installation would need the same private key, which is a non-starter. ABCA is the right answer: the vendor's backend acts as an attester, each installation generates its own key, and the attester signs a short-lived attestation for that key.

The missing link is the AS learning "this vendor's backend is the attester for this client." Without that, every AS has to be pre-configured for every CIMD client's attester, which defeats the point of CIMD. This draft puts that link in the one place the AS is already reading.

The draft is also right that the client's own jwks_uri is a bad place for attester keys. Section 3 explicitly forbids crossing the streams: endorsed keys MUST NOT verify client authentication assertions, and the client's own keys MUST NOT verify attestations. ABCA 10.8's third bullet suggests the opposite, and that suggestion should probably come out of ABCA regardless of what happens to this draft.

The honest problem with publisher-authorized key selection

Section 7.2 says it plainly: under publisher-authorized key selection, "a Client Attestation verified under this policy carries no assurance independent of the publisher." The publisher can run its own attester and claim whatever it wants about platform or hardware integrity.

That's fine, and it's the realistic mode for open CIMD ecosystems. But it means the attestation is doing a different job than the word "attestation" implies. It's proving "the publisher's backend issued a key to this instance," not "this instance is running on a genuine device." That's still valuable, since it's what gets you per-instance keys without a shared secret. The spec should lead with that framing so deployers don't overread it.

Should this be folded into ABCA?

The core of it, yes.

The wire-visible part of this draft is tiny. It's one metadata parameter with two members, plus the rule that iss and kid are present in the attestation. That's half a page. It fits naturally next to the existing client metadata in ABCA Section 9 and would replace the vague third bullet in Section 10.8 with something concrete.

Folding it in would also fix a real awkwardness. ABCA lists "kid plus client metadata" as a recommended key resolution approach, and this draft has to spend a paragraph explaining why that recommendation doesn't actually work. Two documents that disagree about the same key resolution path is worse than one document that gets it right.

The counterargument is timing. ABCA is at -12 and the WG has been trying to close it out. Adding a new metadata parameter now reopens review. If that's the WG's call, a standalone document is fine, but it should be cut down to roughly the half page that matters. Most of the remaining 28 pages is AS policy that doesn't cross the wire.

Significant interoperability concerns

The client can't tell whether its endorsements are enforced. Section 2.3 says the presence of client_attesters doesn't determine whether the profile applies, and publishing it doesn't require the AS to apply it. The AS metadata boolean applies to the AS as a whole, so deployments "establish through a trust agreement" that the profile applies to them. The introduction sells the ability for a client publisher to "always withdraw or narrow its endorsements." That's only true at an AS that has quietly agreed to enforce them. For a CIMD client meeting a new AS, there is no trust agreement. A simpler rule would work better: if the AS advertises support and the client publishes client_attesters, the AS enforces it.

Two key-trust policies with issuer-wide precedence. If the AS has configured trust for an issuer string for any client, that configuration governs that issuer for every client, overriding publisher selection. Removing a configured entry doesn't fall back to publisher selection either. These rules are sensible inside one AS, but they're invisible to clients, and a client publisher has no way to predict which policy will apply to its endorsement.

Withdrawal latency is unbounded. The AS picks the cache maximum age. The spec gives one hour as an example and sets no upper limit. Since withdrawal is the main thing the client publisher gets out of this, it's worth considering a ceiling or at least a SHOULD.

client_id becomes mandatory on every request. ABCA only compares client_id with sub when client_id is present. This profile requires it so the AS can pick the metadata source before evaluating the attestation. That's a reasonable narrowing, but it's a behavior change from ABCA that client libraries have to know about.

Mechanisms that could be removed

Configured aliases (Section 5.3). Aliases exist so the AS can accept an endorsed jwks_uri that differs from its configured source. Under AS-configured trust the endorsed URL is never fetched, only compared. The comparison catches a misconfiguration, but the alias machinery, with issuer-wide scope and tenant-preservation rules, is a lot of normative text for an AS-internal equivalence table. A sentence saying "the AS MAY treat other URIs as equivalent by local configuration" would cover it.

client_attester_endorsement_supported. As written, it tells the client nothing actionable, since Section 2.3 says it doesn't mean the profile applies. Either give it teeth or drop it.

Error-code minutiae (Section 5.4). Two pages on which error code to return in which mode, including which specific ABCA errors leak that endorsement validation succeeded. The net rule is "use invalid_client_attestation." The rest could be a short paragraph.

Dynamic client registration authority rules (Section 2.2). The rule that a registration access token alone can't set client_attesters is a good security point. The paragraph on how to handle RFC 7592 omission-as-deletion semantics reads like an implementation note. And Section 7.6 then points out that the same registration access token can change token_endpoint_auth_method or jwks anyway, which undercuts the protection.

Writing quality

Precise and careful, in the same style as the author's other recent drafts. The two-authority framing is clear and repeated consistently. The security considerations are honest, especially 7.2 on publisher-selected attesters and 7.7 on unscoped endorsements.

It's also very dense. Section 5.3 is a single run of conditional sentences covering policy precedence, aliasing, tenant scope, string comparison, cache keying, JOSE header restrictions, and key eligibility. Much of the document is spent explaining what the endorsement doesn't do: it doesn't make the attester trusted, doesn't establish user delegation, doesn't affect resource servers, doesn't show up in tokens, doesn't change introspection. Those are all true, and a few of them matter. Most could go.

29 pages for one metadata parameter is a lot. The examples in Appendix A and B are good and show the full flow end to end.

Background of the author

Karl McGuinness ([email protected]), Independent

  • 0 published RFCs
  • 16 active Internet-Drafts and 2 expired. Not new to the IETF, and writing at a very high rate.
  • Editor on ID-JAG, an adopted WG document. Also the author of the Identity Continuation Assertion, reviewed here last month.

Questions worth raising

  1. Can the parameter go into ABCA? The wire format is half a page. If ABCA can absorb it before WGLC, the ecosystem gets one document instead of two that disagree about how jwks_uri relates to attester keys.
  2. Should publishing client_attesters mean something? If an AS advertises support, should it be required to enforce a client's published endorsements? Without that, the withdrawal story only works under a private agreement.
  3. How do CIMD clients describe the assurance level of their attester? Publisher-selected attesters carry no assurance beyond the publisher. Is that the expected default for open ecosystems, and should it be stated up front?
  4. What's the minimum version? Cutting aliases, the AS boolean, and most of the error-code text would leave a short, implementable profile.

For the base mechanism this builds on, read Attestation-Based Client Authentication, Section 10.8 in particular.