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

Review · November 19, 2026

Security Principal and Verifier Binding for Agent Communication Protocols

draft-bu-agentproto-security-principal-binding -08View on Datatracker →

The Problem

Agent protocol drafts tend to let one artifact stand in for several security properties. A session ID gets read as delegated authority. An audit receipt gets read as proof the action was authorized. A valid organizational credential gets read as proof of which live agent instance is acting.

This draft does not define a protocol. It defines a review template. For every security claim a protocol carries, the protocol's authors fill in a row: which field carries it, who verifies it, what it is bound to, how freshness works, what happens on failure, and what the verifier is allowed to hand to the application.

How It Fits In

There is no wire format, no token and no IANA request. The only normative references are the BCP 14 keyword RFCs. OAuth, GNAP, WIMSE, RATS and SCITT appear only as examples of things a row could point at.

So does it fit the AGENTPROTO charter? Only loosely.

The charter has three deliverables: a dialog management protocol, a reference architecture, and use cases and requirements. This draft is none of them. The closest home is the reference architecture, which is supposed to describe the identity, authentication and authorization building blocks the protocol reuses. A short claim-separation section could feed into that.

A large part of the draft is about evidence records, receipts and conformance reporting. The charter lists "audit data structures and their production, validation, or consumption" as out of scope.

It does not establish new relationships between OAuth roles. Its vocabulary is closer to RATS: claims, evidence, appraisal, relying-party decision.

Summary

The draft has four layers.

A claim registry. Fourteen provisional claim classes, C-001 to C-014. These cover instance identity, human or organizational authority, delegated scope, session continuity, action evidence, tool identity, freshness, failure handling, standing grants and scoped human continuity.

A verifier matrix. Each row has about 20 fields. The key ones are carrier, verifier, binding, freshness, failure behavior and "accepted result". The accepted result is the constrained output the application may consume, such as "holder of enrolled key on this connection". It is deliberately not the raw token.

Status and evidence labels. A row is specified, planned, inherited or an assumption. Separately it is implemented, partial, none or external. Implementation claims need a pinned revision and at least one rejected negative case.

Row outcomes and composition. Evaluating a row yields one of six outcomes: satisfied, unsatisfied, indeterminate, unsupported, not-evaluated or not-applicable. Rows can depend on other rows. A failure anywhere in the dependency chain propagates, and an aggregate result can never be stronger than its inputs.

The rest is examples: five reusable row patterns, 62 negative test cases, a JSON sketch of a machine-readable mapping set, and a section applying the matrix to other people's drafts.

The core ideas are sound. For OAuth readers, most of them are familiar: a sender-constrained token proves possession and not scope, a token is not a decision, and offline signature validation is not a revocation check.

Interoperability

Significant interoperability concerns

There is nothing to interoperate with. The draft says the claim registry is "not a conformance target", the JSON field names are not normative, and the row can be a table, YAML, a Markdown issue or prose. Two drafts can both "use the matrix" and produce documents that cannot be compared mechanically.

It still uses MUST about 50 times. Section 16 says a row result "MUST identify its stable result identifier, mapping-set identifier and revision, row identifier, claim identifier, outcome, verifier and verifier-profile revision, and evaluation context". No format carries those fields. These are requirements on a data structure the draft declines to define.

The claim registry is unstable. The classes overlap. C-008 (freshness), C-009 (failure handling) and C-011 (accepted result) are also columns in every row. Logical-agent continuity is described at length in Section 8 but has no claim ID. The open issues list asks what the minimum mandatory set is.

The vocabulary is used inconsistently in the draft's own examples. Sections 24.3 and 24.5 give "Specification status: external". External is an implementation-status value. The specification-status value for this case is "inherited". Section 32 refers to "thirteen claim classes" when the registry has fourteen.

Mechanisms that could be removed

  • The evidence campaign rules in Section 15. Amendment policies, allowance slots, ordinals and the provenance type of an allowance limit are a methodology for honest test reporting. They have nothing to do with agent protocols.
  • Evaluation-plan pinning in Section 16. The same test-reporting concern, applied to check ordering.
  • The check-class registry. A versioned, extensible registry with collision analysis, inside a document that makes no IANA request.
  • The machine-readable mapping set. The draft says it is not normative. It belongs in a repository.
  • Section 32's test results. It reports that 13 of 22 test vectors failed against a pinned commit of another author's IACP prototype, with SHA-256 hashes. That is a mailing list post or a GitHub issue. It will be stale before the draft expires.
  • Most of the 62 negative cases. Many restate a matrix rule as a scenario.

The matrix fields, the status values and the "accepted result" concept would fit in ten pages.

Writing Quality

The grammar is clean and the document structure is conventional. The problem is length and repetition.

This is 74 pages with one author. The file grew from 3,584 lines at -06 to 4,144 at -08. The row fields are listed three times, in Sections 10, 13 and 18, and the three lists do not match. The same sentence shape, "X does not by itself prove Y", carries most of the document.

Each revision appears to absorb the latest mailing list thread. The acknowledgments run two pages and thank people for individual distinctions, which map one-to-one to added paragraphs. That is how a review template ended up with rules like "Two distinct amendments MUST NOT consume the same allowance slot".

The later sections are hard to read. Section 15's paragraph on amendment allowances is 30 lines of nested conditions with no example.

Some justification is missing where it matters most. The draft never shows one complete filled-in row for a real protocol. The "worked example" in Section 25 lists possible carriers for each claim and stops there.

Authors

From the Datatracker profile:

  • Songbo Bu. No affiliation listed. No RFCs. 5 active drafts and no expired ones, so all of this work is recent. The others are in TLS and RATS, plus co-authorship of Agent Accountability: Composition and Conformance, which this draft cites as an early application of the matrix.

Questions for the Working Group

The draft's own first open issue is the right one: is this a requirements document, a security considerations companion, or text to import into other drafts?

  • Does the reference architecture need a claim-separation section? If so, the six or seven core row fields are the part worth taking.
  • Should AGENTPROTO candidate protocols be asked to state, per claim, the verifier and the failure behavior? That is a useful review habit and does not need an RFC.
  • Is anyone other than the author maintaining a mapping? The IACP and AIP mappings in Section 32 are the only evidence of use, and the draft says the AIP experiment does not establish working group agreement.

The use cases and requirements draft is the adopted document where security requirements for the dialog protocol are being collected. Feedback that the WG wants claim separation reflected there is likely to have more effect than another revision of this one.