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

Review · November 21, 2026

Zero Trust Fabric Layer Agent-to-Agent Chained Trust on a Connected Flight

draft-seymour-wimse-connected-flight -06View on Datatracker →

The problem

An agent hands part of its task to a second agent, which hands part to a third. What stops a hop in the middle from quietly reaching into another tenant?

The draft says nothing does today. Token Exchange records prior actors in nested act claims, and RFC 8693 tells the consumer to ignore them for access control. A child token forwarded down the chain can be replayed at any hop.

The draft proposes two checks. Every hop must trace back to one credential issued when the task started. A hop that crosses a tenant boundary needs an extra credential issued by the destination tenant.

How it fits in

The draft is addressed to WIMSE, and multi-hop chains across trust domains are in the WIMSE charter. The draft doesn't build on the WIMSE work though. It never mentions the Workload Identity Token, the Workload Proof Token or the WIMSE architecture. AIMS appears only as related work with a gap to close.

The real base is the Zero Trust Fabric Layer (ZTFL), an SSRN paper by the same author. ZTFL is described as seven enforcement points built on AWS IAM and VPC Lattice. Four of the six terms in the draft's evaluation function are ZTFL conditions that this document never defines.

Token Exchange and Mutual-TLS are normative references. Token Exchange is cited mostly to be argued against. Mutual-TLS supplies the certificate thumbprint that binds a Visa to a connection. RATS and EAT are normative for one optional section.

The draft also sits in a cluster of individual drafts that cite each other. It maps itself onto the verifier evaluation draft, and defers to three Schrock drafts for action identifiers and single-use consumption. None of these are adopted anywhere.

No OAuth role is reused. The Identity Broker, the STS, the boundary issuer and the key registry are all new parties with no stated mapping to an authorization server or resource server.

What it specifies

The model is four credentials, named after international travel:

  1. Passport. The agent's long-lived identity, the same at every hop. Issued by an Identity Broker.
  2. Ticket (Root Token). Issued once by the STS when the task starts. It carries the approved action (the "frozen call") and a hard maximum lifetime.
  3. Boarding Pass. A short-lived token issued by the STS for one hop. It names the Ticket and a hop index. A Boarding Pass that can't be traced to the Ticket is rejected.
  4. Visa. Needed only when a hop targets a resource in another tenant. It is signed by the destination tenant's issuer key, scoped to that tenant, and bound to the caller's mTLS certificate.

Most hops stay inside one tenant and never need a Visa. That is the efficiency argument: only the boundary-crossing hop pays for the extra check.

The draft claims this resists prompt injection. A redirected agent can ask for the next hop, but it can't mint its own Boarding Pass.

A Cedar schema and two policies encode the checks. Five test requests run in the Cedar playground produce the expected allow and deny results.

An optional section lets Passport issuance depend on a RATS hardware attestation result.

Is this enough for interoperability?

No. This is an architecture description with illustrative JSON, and the draft says so: the field names are "illustrative, not a registered claim set."

No token format. No signature format, no claim definitions, no key discovery. The Ticket example has no issuer. The Boarding Pass "embeds a signed reference" to the Ticket, and the draft never says what that is.

No issuance protocol. How does an agent request a Boarding Pass? How does a hop in tenant A get a Visa from tenant B? Neither request is defined. The draft lists how Visa issuance is authorized as future work.

The Boarding Pass has no audience. It names a holder and a hop index. Nothing names the resource or the next agent. The draft never says how a verifier learns the current hop index either.

The efficiency claim doesn't hold. The draft criticizes Token Exchange for a round trip to the authorization server at every hop. Its own Boarding Pass is issued by the STS at every hop. Section 7 even says Token Exchange "may still issue the underlying Boarding Pass." The comparison table lists the within-tenant cost as "none."

Local verification contradicts itself. Section 5 says verification "never requires a synchronous round trip to STS." Section 4.1 has the verifier check issuer standing against the STS's current registration at evaluation time. Ticket teardown on chain completion is also STS state that a local verifier never sees.

The standing source comes from the token. The new issuer_standing_ref URL sits inside the Boarding Pass. The verifier evaluation draft requires that source to be pinned by the relying party, not taken from the presented token. An attacker who forges a Boarding Pass also picks the URL.

The prompt injection claim is weaker than stated. An injected agent doesn't need to forge anything. It asks the STS for the next Boarding Pass like any other hop would. Whether the STS says yes depends on issuance policy, which is not specified.

Single use is admitted as missing. A valid Boarding Pass can be presented more than once before it expires. The draft says this plainly and promises a mechanism in a future revision.

The Cedar test proves little. The policies compare strings in a request context the caller supplies. There are no signatures, no mTLS binding, no Visa issuer check and no Ticket lifetime check. Test 3, the "forged" Boarding Pass, is a string mismatch.

The Visa registry is described three ways. Section 4 calls it the destination tenant's own key set. Section 9 recommends a shared registry and talks about keys for the "sending boundary." Section 4.4 says Section 9 defines the registry, and it doesn't.

The intended status is Informational, yet the text uses MUST for token lifetimes and key separation.

What could be removed

Section 4.1 maps the model onto another individual draft's "five inputs and four rules." That draft now lists six inputs, so the mapping is already out of date. It belongs in a mailing list thread.

Section 4.2 on hardware attestation and RF-PUF is unrelated to chaining. RF-PUF has no reference.

Section 4.3 describes a consumption mechanism the draft doesn't define.

The ZTFL background, the Delegation Logic and Zanzibar comparisons, and the Cedar test table are academic paper material. The embassy in Kingston passage and the airport diagram restate the analogy without adding a requirement.

What's left is two requirements. Every hop token carries a verifiable reference to the transaction's origin. A cross-domain hop needs a grant issued by the destination domain. Both are reasonable inputs to the WIMSE token exchange work and to transaction token discussions, as requirements on those formats.

Writing quality

The prose is fluent and grammatical. It reads like a journal article: related work, formal model, validation, comparison table, future work.

It is not IETF style. Requirements are buried in narrative. The analogy does the work that definitions should, and the terms drift. The same credential is called Ticket, Root Token, Root Ticket and K. The Boarding Pass issuer is the STS in one place and "the carrier of record" in another.

Revision history is written into the body: "This revision further shows", "new in this revision", "This document previously left open."

At 26 pages it is long for two requirements. The traceability rule is restated in at least six sections.

One claim needs support. Section 8 says the traceability requirement "informally shaped" the delegation chain draft, which "adopts the same requirement." The current revision of that draft doesn't cite this document.

The references need cleanup. The arXiv reference has no authors, and the verifier evaluation reference points at -03 when that draft is now at -05.

Author

Errol Seymour has no RFCs. This is the only document on the Datatracker profile, with no expired drafts. The draft lists Capitol Technology University as the affiliation. This is a new IETF author.

The draft is moving fast. Seven revisions were posted between 18 September and 1 October, two of them on the same day. Nothing has been posted since.

Bottom line

So what should the working group do with this? Not adopt it. There is nothing here a second implementer could build against, and the base it extends is an AWS-specific architecture from a single-author paper.

The core observation is sound. A chain should be tied to its origin, and the destination domain should have a say when a chain crosses into it. The useful next step is to bring those two points to the WIMSE list as requirements on the WIMSE token exchange profiles and on AIMS, and to drop the parallel credential system.