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

Review · November 16, 2026

WIMSE Workload Identity Metadata

draft-heigh-wimse-workload-identity-metadata -00View on Datatracker →

The Problem

Deployments stuff attributes into workload identifiers. The draft's own example is spiffe://example.com/ns/prod/team/payments/role/reconciler/service/reconciliation. Relying parties then parse the path and authorize on the pieces.

The draft says to stop doing that. The identifier names the workload. Group, role, and entitlements travel separately as Workload Metadata, either as claims in a JWT credential or from an authenticated lookup API.

How It Fits In

This is an individual draft at -00 with a 10 minute slot on the IETF 127 WIMSE agenda.

It builds directly on the adopted WIMSE documents. Workload Identifier already tells consumers to compare the complete URI and warns against prefix matching. Workload Credentials already allows extra registered claims in a WIT.

It adds one new party. Section 7 says the metadata service can be run by the credential issuer or by "a separate authority".

New registrations: one JWT claim (metadata_uri), one X.509 certificate extension, and one well-known URI (workload-metadata). The groups, roles, and entitlements claims are reused from RFC 9068 and SCIM (RFC 7643).

A note on the name. This is not metadata in the RFC 8414 or RFC 9728 sense. Those are public, cacheable documents describing a server. This is an authenticated, per-subject attribute lookup with per-caller filtering. It is closer to OpenID Connect UserInfo for workloads.

Summary

The draft does three separate things.

A principle (Sections 3 and 4). Attributes do not belong in identifiers. Identity credentials SHOULD NOT carry scopes or fine-grained permissions (Section 4.3). Those belong to an authorization server.

In-band claims (Sections 4.1 and 6.1). A WIT or JWT-SVID carries groups, roles, and entitlements. A relying party MUST read them as describing the workload, never an associated person.

A lookup API (Section 7). Workload B authenticates to the metadata endpoint and sends GET ?sub=<workload A's identifier>. The service returns a JSON object with sub plus whatever claims B is allowed to see.

How does the client find the endpoint? Section 7.1 gives a strict order: local configuration, then a URI carried in the validated credential, then https://<trust-domain>/.well-known/workload-metadata. There is no fallback from one to the next on failure.

X.509 credentials carry no attributes at all. They get only a pointer, a new non-critical extension holding the endpoint URI (Section 6.2.1).

Does this overlap with Trust Domain Discovery?

Partly. The blog reviewed that draft in August.

The function is different. Trust Domain Discovery publishes keys, unauthenticated, to anyone. This draft returns attributes, authenticated, to authorized callers.

The plumbing overlaps. Both drafts register a well-known URI on the trust domain's host: wimse-trust-domain there, workload-metadata here. Section 5.1 of that draft says future specs may register additional members. A workload_metadata_endpoint member there would replace the default URI here.

This draft does not cite Trust Domain Discovery. Yaroslav Rosomakho is an author of both.

Interoperability

Significant interoperability concerns

No mandatory caller authentication. Section 7.2 lists mutual TLS, HTTP Message Signatures, or WIT plus WPT, and picks none. A client that only does mutual TLS cannot query a service that only accepts WPT. The draft also does not say how the service comes to trust a caller from another trust domain.

The claim encoding is unsettled. Section 4.1 says implementations MUST use the registered claims with "the same meanings and value encodings". The same section then says SHOULD emit arrays of strings, which is not the object form RFC 9068 points to. Open Issue 2 admits the conflict. Values have no shared vocabulary either.

No caching at all. Section 7.5 requires Cache-Control: no-store on every response. Clients MUST NOT reuse retrieved metadata for any other authorization decision. That reads as one lookup per decision. Section 8 then says access MUST NOT be granted if required metadata cannot be fetched. The metadata service becomes a hard dependency on every request. Open Issue 4 covers this.

The well-known default does not remove configuration. Section 7.1 says the client's trust configuration MUST authorize the mapping before the client contacts the endpoint.

Mechanisms that could be removed

  • The well-known URI, in favor of a member in the Trust Domain Metadata document.
  • The platform credential rules in Section 4.4 and the issuance rules in Section 11.2. These are MUST-level rules about issuance, not metadata. They fit better in the identifier draft or Workload Identity Practices.
  • The X.509 extension could be its own document. It needs a PKIX OID and will want review from certificate people.

Writing Quality

The writing is very good for a -00. The endpoint rules are precise: exact sub matching, omission is not a negative assertion, 404 for both unknown and unauthorized subjects.

It is long. The text is about 1,900 lines, and roughly a quarter of it is argument for the separation principle.

Some of the quoted text is already stale. The draft cites workload-creds-02 and identifier-03. Both were revised after this draft was posted.

  • Sections 3.1 and 6.2 quote Workload Credentials as saying correlation data "MUST NOT be encoded as a second workload identifier". That sentence is not in -03.
  • Section 4.4 quotes the identifier draft as saying issuers "MUST ensure uniqueness of all Workload Identifiers they assign". The -04 text is reworded.
  • The AIMS reference is to -00. The current version is -01.

Authors

From their Datatracker profiles:

  • Shlomo Zalman Heigh (Palo Alto Networks). No published RFCs. This is his only draft.
  • Joseph A. Salowey (Palo Alto Networks). 23 RFCs. 6 active drafts, 5 of them adopted WIMSE documents including the architecture, identifier, and credentials drafts. Chair of TLS, EMU, Privacy Pass, and MOLE.
  • Yaroslav Rosomakho (Zscaler). 1 RFC, RFC 9850. 28 active drafts across WIMSE, TLS, MASQUE, QUIC, and OAuth. IAB member. Chair of AUDIT, HPKE, and SEAT.

Questions for the Working Group

So what can a 10 minute slot settle?

  • Is this one document or three? The separation principle could be a paragraph in the identifier draft. The claims profile is small. The API is the large new piece.
  • Does WIMSE want an online attribute lookup service at all?
  • Should endpoint discovery go through the Trust Domain Metadata document instead of a second well-known URI?
  • Which caller authentication binding is mandatory to implement?
  • Who decides the encoding of groups, roles, and entitlements, WIMSE or OAuth? Open Issue 2 suggests raising it against RFC 9068.
  • Is no-store with no reuse workable, or does the draft need a bounded cache lifetime before adoption?