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

Update review · November 20, 2026

WIMSE Applicability for AI Agents -04, Shorter and Cleaner but Still No Credential Format

draft-ni-wimse-ai-agent-identity -04View on Datatracker →

The -04 is a cleanup. The authors cut two sections, fixed the grammar, and added two co-authors. The draft's own changelog says "Fixed editorial issues, and streamlined the content," and that's accurate.

The core problems from our review of -02 are still open. The dual-identity credential still has no format, and Security Considerations still says "TODO Security."

What changed

  • Two new co-authors. Michael Richardson (Sandelman Software Works) and Bingzheng Wang (CESI) join Yuan Ni and Chunchi Peter Liu.
  • The framing moved from delegation to accountability. -02 said the agent "operates on behalf of" its owner. -04 says the owner is "accountable for the operation of" the agent. The credential is now described as a foundation for "accountability and audit," where -02 said "access control and accountability."
  • The agent now signs its own request. In the agent-mediated and owner-mediated models, the agent "generates and signs" the request, the owner countersigns it, and the server "verifies both signatures." In -02 only the owner signed.
  • The Identity Proxy definition is narrower. It's now "an intermediary component between an agent and the Identity Server." The -02 text said it could "request, inspect, replace or augment" credentials.
  • The campus network use case has its own section number. It was a broken markdown heading in -02.
  • The WIMSE architecture reference is current. It now points at -08.

The page count went from 14 to 13.

Was anything added?

One thing. A new Section 4.3, "Hardware Root of Trust," is two sentences long. It says the owner can sign with a hardware root of trust, so multiple virtual agent identities can be derived from one device.

That's a statement of possibility. It doesn't say how the derivation works or how a verifier would know the owner key is hardware-backed.

What was removed?

  • The CHEQ comparison. We suggested cutting it. It's gone, along with the reference.
  • The entire Initial Trust Establishment section. The BRSKI and EST discussion is deleted, and RFC 8995 and RFC 7030 are no longer referenced.
  • The EAT profile for AI agents. Section 3.2 no longer points at draft-messous-eat-ai for the structure of attestation evidence. The evidence format is now unspecified.
  • The CA assumption. -02 presumed the identity server held a signing certificate with keyCertSign set. That sentence is gone.
  • The Human-in-the-Loop claim. -02 called the dual-identity credential "a manifestation of Human-in-the-Loop mechanisms." -04 drops it.
  • A security mitigation. -02 said mutual authentication and cryptographic integrity were "mandatory between proxies and the owner" in gateway mode. -04 deletes that sentence and keeps only the rate limiting advice.
  • The A2A reference.

The removed keyCertSign sentence was the only hint that the credential is an X.509 certificate. Without it, the draft says nothing about what kind of credential this is.

Was the previous feedback addressed?

Partly. Here's where each point stands.

  • Three issuance models with no selection guidance. Not addressed. All three are still there, and nothing says which one to implement or how the parties know which one is in use.
  • No credential format. Not addressed. There is still no wire format, no JWT or X.509 mapping, and no mention of how this relates to the WIMSE Workload Identity Token or workload certificates.
  • Security Considerations is "TODO." Not addressed. Acknowledgments is also still "TODO acknowledge."
  • The open question on bootstrapping. Resolved by deletion. The question is gone because the section is gone.
  • Cut the CHEQ comparison. Done.
  • Trim the campus 802.1X use case. Not done. It was promoted to a numbered subsection and got slightly longer.
  • No normative language. Not addressed. The BCP 14 boilerplate is still in Section 2 and the document still contains zero uppercase keywords. RFC 2119 and RFC 8174 are now the only normative references.
  • Writing quality. Addressed. The typos and awkward phrasings we called out are fixed. A stray "Document History" list is sitting inside IANA Considerations, which is minor.

So what does the BRSKI removal mean?

It's the odd part of this revision. Michael Richardson co-authored BRSKI (RFC 8995), and the BRSKI section was deleted in the same revision that added them as an author.

The likely reading is that bootstrapping is being scoped out so this draft only covers owner binding. That's a reasonable call. The draft should say so, because step 1 of the architecture still assumes the proxy and server "have established a secure channel" with no pointer to how.

The accountability reframing is the more useful change. "On behalf of" invited the question of whether the agent inherits the owner's permissions. "Accountable for" doesn't. The campus use case undercuts it though. It still has the network assigning the agent to Alice's R&D VLAN based on her existing RBAC roles, which is permission inheritance.

Background of the Authors

  • Michael Richardson (Sandelman Software Works, new): 29 RFCs, including BRSKI (RFC 8995), the voucher format (RFC 8366) and the RATS Architecture (RFC 9334). They have 31 active drafts and 73 expired ones, chair the ASDF working group, and serve as secretary for ANIMA and ROLL. A very experienced author who is writing a lot right now, mostly in device onboarding and attestation.
  • Bingzheng Wang (CESI, new): 0 RFCs, no expired drafts, and one active draft, which is this one. A new IETF author.

Yuan Ni and Chunchi Peter Liu (both Huawei) were covered in the -02 review.

Richardson brings exactly the background the hardware root of trust and attestation parts need. Those are currently the thinnest sections in the document.

What to do with this version

Skim Section 4.2 if you want to see the countersigning change. The rest reads the same as -02 with better English.

The feedback to send is the same as last time, in this order:

  1. Define the dual-identity credential. Say whether it's an X.509 certificate, a WIT with an owner claim, or something new, and show an example.
  2. Pick one issuance model, or say how a deployment selects one.
  3. Write the Security Considerations. The attack surface notes under each model are a start, and the deleted mutual authentication requirement belongs there.

The credential format question is best asked against the WIMSE architecture, since that's the document this one claims to apply.