Review · November 16, 2026
OAuth SPIFFE Client Authentication
The Problem
A workload running under SPIFFE already has an attested credential, its SVID. To call an OAuth token endpoint today it still needs a second credential, a client secret or a private_key_jwt key, provisioned out of band.
This draft lets the workload use the SVID itself as its OAuth client credential. It covers all three SVID formats: JWT-SVID, X509-SVID, and WIT-SVID.
How It Fits In
This is an adopted OAuth working group draft, not a new submission. It is at -03 and is on the joint OAuth/WIMSE session agenda at IETF 127.
It defines no new roles. It adds three client authentication methods (spiffe_jwt, spiffe_x509, spiffe_wit), one client assertion type, and two client metadata parameters (spiffe_id, spiffe_bundle_endpoint).
Each method profiles an existing mechanism:
- JWT-SVID profiles RFC 7523 client assertions.
- X509-SVID profiles RFC 8705 mutual TLS.
- WIT-SVID profiles Attestation-Based Client Authentication, using the WIT from WIMSE Workload Credentials.
So what is left before publication? Mostly dependencies. The draft has four normative references to unfinished IETF drafts: attestation-based client auth, rfc7523bis, WIMSE workload credentials, and Client ID Metadata Document. The attestation and WIMSE references were informative in -02 and became normative in -03. The SPIFFE references are undated GitHub markdown files. The WIT-SVID one points to a branch named draft-wit-svid, not main.
The JWT-SVID and X509-SVID parts depend on much less. The WIT-SVID part is what ties this document to the attestation and WIMSE drafts.
Summary
The authorization server is configured with a SPIFFE trust domain and a way to get its keys. The client presents an SVID at the token endpoint. The server validates it against that trust domain's bundle and maps the SPIFFE ID to a client.
- JWT-SVID (Section 3.2): sent as
client_assertionwith the new typeurn:ietf:params:oauth:client-assertion-type:jwt-spiffe. Theaudmust be the authorization server's issuer identifier as its sole value. - X509-SVID (Section 3.3): the SVID is the mutual TLS client certificate. Path validation MUST use only that trust domain's anchors, never the system trust store.
- WIT-SVID (Section 3.4): the WIT goes in the
OAuth-Client-Attestationheader withtypofwit+jwt. The client signs a PoP JWT with the key bound in the WIT'scnf.
Keys come from a SPIFFE Bundle Endpoint using the https_web profile (Section 6.2). The endpoint URL cannot be derived from an SVID and has to be configured, or published in a Client ID Metadata Document.
What changed in -03
This is a large revision. The -02 text was about 1,400 lines and -03 is about 1,680.
client_idis now OPTIONAL for all three methods, with one set of identification rules in Section 3.1. Without it, the SPIFFE ID is the client identifier. For X509-SVID this deviates from RFC 8705 Section 2, and -02 required the parameter.- New Sections 6.1 and 8.2 cover trust bundles fetched from a Client ID Metadata Document.
- New Section 8.1 covers JWT-SVID replay.
- The three method names are now registered in the Token Endpoint Authentication Methods registry.
- The -02 TODO about
attest_jwt_client_authis resolved. That value does not imply WIT-SVID support. - X509-SVID validation now references the SPIFFE spec instead of restating it.
Interoperability
Significant interoperability concerns
There is no mandatory method. Section 4 says servers and clients each MUST support "at least one" of the three. A JWT-SVID-only client and an X509-only server are both conformant and cannot talk. The fix offered is that clients SHOULD support all three. The one listed implementation, Keycloak in preview, covers JWT-SVID only.
The only implemented method is the weakest. Section 8.1 says it plainly: "An attacker that obtains a JWT-SVID can use it for client authentication until it expires." SPIFFE does not require jti, and implementations may hand the same cached JWT-SVID to a workload repeatedly. The audience check limits the damage to one authorization server.
Bundles from a Client ID Metadata Document change what the SPIFFE ID means. Section 8.2 is honest about this. Whoever controls the client_id URL picks the keys. A document at attacker.example can declare spiffe://bank.example/payments with its own bundle. The server MUST NOT treat that SPIFFE ID as issued by the trust domain. The five mitigations are a SHOULD for "one or more". Two servers can make different choices and accept different clients.
At that point the mechanism is close to private_key_jwt with a jwks_uri in a different key format. The working group should ask whether that is worth a new metadata parameter and two pages of security considerations.
Refresh tokens under WIT-SVID are not addressed. The attestation draft (Section 10.3) binds refresh tokens to the Client Instance Key unless a profile says otherwise. Section 3.3 explains what rotation does to certificate-bound access tokens. There is no equivalent text for WIT-SVID. If the workload's key changes when its WIT-SVID is reissued, the refresh token stops working.
iss-based key discovery is still allowed. Section 6.3.4 permits it as a "compatibility mechanism". Section 8.3 then needs four MUST-level rules to make it safe.
Mechanisms that could be removed
spiffe_bundle_endpointin Client ID Metadata Documents, for the reason above.- The
issfallback in Section 6.3.4. - WIT-SVID could move to a separate document. It carries the unfinished dependencies and has no listed implementation.
Writing Quality
The writing is clear and uses standard IETF language. The numbered validation steps are easy to implement from. Section 8.2 is a good example of explaining a trust model instead of hiding it.
The -03 restructuring left some loose ends:
- Section 6 says it "recommends two SPIFFE-native approaches", then gives an example using the Workload API. Section 6.3.1 says the Workload API is NOT RECOMMENDED.
- Sections 6.1 and 8.2 refer to bundles "obtained as described in Section 6". Section 6.1 is part of Section 6.
- Section 6.3 says its methods "SHOULD be avoided". One of them is a MUST NOT and one is a MAY.
- Section 4 uses a lowercase "should" for the revocation and introspection metadata.
spiffe_idis marked REQUIRED without saying when.- The reference to Client ID Metadata Document cites -02. The current version is -03.
- The author block spells "Kasselmann". Datatracker and his other drafts say Kasselman.
Authors
From their Datatracker profiles:
- Arndt Schwenkschuster (Defakto Security). No published RFCs. 8 active drafts, including WIMSE Workload Credentials and OAuth Identity Chaining. A very active writer across OAuth and WIMSE.
- Pieter Kasselman (Defakto Security). 1 RFC, RFC 10027 on cross-device flow security. 7 active drafts. WIMSE chair.
- Scott Rose (NIST). 10 RFCs, nearly all DNS and DNSSEC. This is his only active draft.
- Stian Thorgersen (IBM). No RFCs. This is his only active draft. The draft lists Keycloak as a CNCF / IBM implementation.
- Nancy Cam-Winget (Cisco Systems). 14 RFCs. 2 active drafts. Chair of SCIM and SEAT.
Questions for the Working Group
So what should the joint session decide?
- Should one method be mandatory to implement? If so, which one?
- Should WIT-SVID be split out so JWT-SVID and X509-SVID can go to last call without waiting on attestation-based client auth and WIMSE credentials?
- Is fetching trust bundles from a Client ID Metadata Document worth keeping, given what Section 8.2 has to say about it?
- What are refresh tokens bound to under WIT-SVID?
- How will the SPIFFE references be made stable enough for a Standards Track RFC? WIMSE and SPIFFE people are in the room for this one.