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

Review · November 16, 2026

WIMSE Architecture

draft-ietf-wimse-arch -08View on Datatracker →

The Problem

Every WIMSE protocol draft needs the same vocabulary: workload, trust domain, workload identifier, workload identity credential, security context. This document is where those terms are defined.

It also describes the components around a workload and the scenarios the protocol drafts are meant to serve.

How It Fits In

This is an adopted WIMSE working group draft at -08, not a new submission. It is Informational, has no IANA actions, and defines no protocol. It has a 10 minute slot on the IETF 127 WIMSE agenda.

It is the first item in the charter's program of work. The charter asks for a document that "defines common terminology, discusses workload attestation and identity, specifies a threat model, and defines a set of architectural components," with 2-3 scenarios and the key interoperability points for each.

Every other adopted WIMSE protocol draft depends on it. Workload Credentials and HTTP Signatures both say "All terminology in this document follows" the architecture. The WPT and mutual TLS drafts cite it too.

So what is left before publication? The text is not finished, and it has fallen behind the drafts that cite it.

  • Section 3.4.10 still contains "(TODO: align with OAuth cross domain identity and authorization)".
  • The Acknowledgments section reads "Todo: Add your name here."
  • The draft cites draft-ietf-wimse-s2s-protocol-07 from October 2025 for the WIT, for trust domain naming, and for both authentication mechanisms (Sections 2 and 3.4.2). That document has since been split into workload-creds, WPT, http-signature, and mutual-tls. The architecture references none of the four.
  • The charter asks for a threat model. Section 4 is six short subsections of general security considerations.

Summary

Section 3.1 names three building blocks: trust domain, workload identifier, and workload identity credential. A credential is either an X.509 workload identity certificate or a Workload Identity Token (WIT). Both bind an identifier to a key that the workload proves possession of.

Section 3.2 gives three scenarios:

  • Basic (3.2.1): a gateway, three workloads, and a CA/credential service in one trust domain. Workloads authenticate each other with mutual TLS or a JWT-based mechanism.
  • Context (3.2.2): the gateway swaps an external OAuth access token for an internal context token, such as a Transaction Token. Workloads pass it along the call chain.
  • Cross-domain (3.2.3): a workload gets a token from its local token service, then exchanges it at an external token service, following OAuth Identity Chaining.

Section 3.3 is new in this revision. It zooms in on one workload-to-workload hop and adds a Policy Enforcement Point and an optional Policy Decision Point.

Section 3.4 lists eleven use cases, from credential bootstrapping to AI agents.

Since -06, workload and workload instance are separate terms, and the identifier definition moved to Workload Identifier.

Does It Match the Other WIMSE Drafts?

Not yet. This is the main finding.

The issuer has different names. The architecture calls it the "CA/Credential Service", and in the same section also "Credentials Service" and "token/Credential service". Workload Credentials Section 5.1 says the WIT "is issued by the Identity Server". The architecture never uses that term.

The WPT and HTTP Message Signatures are missing. These are the two proof of possession mechanisms the working group adopted. The architecture mentions neither by name. Section 3.3 says only "proof of possession, as defined in the credentials and protocol documents."

The certificate is not defined here. "Workload Identity Certificate" appears throughout but is not in Section 2. The mutual TLS draft points to Workload Credentials Section 6.1 for the definition.

Bearer tokens. Section 2 says deployments "may also deploy bearer tokens as workload identity credentials". Workload Credentials Section 9.3 says "The WIT MUST NOT be used as a bearer token." These can both be true, but the architecture should say which credentials it means.

Identifier rules are stated twice. Section 3.1.2 defers to the identifier draft, then restates its authority MUST and FQDN SHOULD anyway. Workload Identifier Section 4.1 now requires a "non-empty" authority and a path. The copy here has already drifted, and the reference is to identifier -03, not -04.

Identity Proxy is used only here. None of the five protocol drafts use the term. An undefined "agent" also "provisions the identity credentials" in Section 3.1.3.

The good news is that the newest pieces line up. The Workload definition is word for word the same in the identifier draft. Section 3.3 matches the message flow in Workload Credentials Section 1.1, which cites it.

Interoperability

The charter asks for the key interoperability points per scenario. The scenarios describe flows and do not list those points.

Normative language without a protocol

The draft uses BCP 14 keywords for things no implementation can be tested against:

  • "Each authenticated request MUST leave a verifiable and inspectable trace" (3.4.5).
  • "each hop in the chain MUST explicitly scope and re-bind the security context" (3.4.11). No mechanism is named.
  • The trust domain to key mapping "MUST be obtained through a secure mechanism" that is out of scope (3.1.1).

Text that could be removed

  • Section 3.4.11 on AI agents. The working group has since adopted AIMS for this. One paragraph pointing there is enough.
  • The duplicated identifier requirements in 3.1.2.

Writing Quality

Readable, with good diagrams. It needs an editing pass before last call.

  • Section 3.2.2 step 1: "the gateway service obtains credentials it can use with the gateway service". This should be the application client.
  • Figure 2 shows a "Context Service". Step 3 calls it "a token service".
  • Section 3.2.3 step 6 requests the access token "from the token service". The figure shows the external token service.
  • Typos: "providioned" (3.4.1), "to associated" (3.2.1).
  • The charter defines a workload as "a running instance of software". The draft now calls that a workload instance. This is the better definition, and the working group should confirm it on purpose.

Authors

From their Datatracker profiles:

  • Joseph A. Salowey (Palo Alto Networks). 23 RFCs, mostly TLS and EAP. 6 active drafts, all WIMSE. Chair of TLS, EMU, and Privacy Pass. Also an author of four of the five protocol drafts.
  • Yaroslav Rosomakho (Zscaler). 1 RFC, RFC 9850. More than 25 active drafts across TLS, MASQUE, and WIMSE. IAB member and chair of SEAT and HPKE.
  • Hannes Tschofenig (University of the Bundeswehr Munich). 99 RFCs. 36 active drafts. OAuth chair.

Questions for the Working Group

So what should the working group decide with its 10 minutes?

  • Does the architecture go to last call before the protocol drafts, or with them? They reference it for terminology, so it has to be stable first.
  • Which name wins: Identity Server or Credential Service?
  • Should the architecture name the WPT and HTTP Message Signatures as its two application-layer proofs?
  • Is a threat model still a deliverable, or is Section 4 enough?
  • Should the AI agent text move to AIMS?
  • Should the TODO in Section 3.4.10 be closed by citing Identity Chaining?