Internet-Draft Approval-Act Record November 2026
Ashcroft, et al. Expires 20 May 2027 [Page]
Workgroup:
Network Working Group
Internet-Draft:
draft-ashcroft-approval-record-00
Published:
Intended Status:
Experimental
Expires:
Authors:
J. Ashcroft
Kipple Labs
B. Morrison
Alter Meridian Pty Ltd
N. Mathews
Agentic Verification Architecture

An Approval-Act Record for Standing Authorization Grants

Abstract

A standing authorization grant is an artifact that outlives the event that produced it. Existing authentication-strength vocabulary (acr/amr) describes authentication events and travels on tokens. Nothing carries a record of how a standing grant's approval was performed to the party verifying the grant, possibly days later and with no human present. This document defines an approval-act record: evidence of the approval ceremony, bound to the exact grant issuance approved, and carried in-band with the grant it authorizes. It defines one initial evidence type (WebAuthn assertions over a hash of the canonical grant), states precisely which of the record's claims a verifier can recompute, which rest on vendor attestation, and which remain attested by the operator, and provides a worked example with a failure matrix.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on 20 May 2027.

▲

Table of Contents

1. Introduction

The Monday/Thursday problem, stated once. An operator approves a delegation on Monday behind a hardware authenticator. The agent acts on Thursday with no operator present. A token minted Thursday carries Thursday's acr: a service credential, or nothing. The approval act that actually stood behind the authority is invisible to every verifier that matters, and it cannot be reconstructed after the fact. If it is not captured when the grant is made, it is gone.

The unattended case is the reason standing grants exist, so this document treats it as the main case.

1.1. Requirements Language

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.

1.2. Terminology

  • Standing authorization grant (grant): a signed artifact delegating authority to an agent, valid across many later presentations.
  • Operator: the party that runs the agent and issues or obtains the grant.
  • Approver: the human whose approval the grant is meant to reflect.
  • Approval act: the ceremony in which the approver approves one grant issuance.
  • Approval record (record): the evidence of the approval act defined in Section 3.
  • Verifier: any party that checks the grant and its record, typically a relying party receiving the agent's request.
  • Operator anchor: the operator's JWK Set, published at the URI the Signature-Agent field names, under a name the operator controls, as used by the Web Bot Auth protocol [I-D.ietf-webbotauth-httpsig-protocol].
  • Covered issued-at: the issuance time carried inside the grant bytes covered by the record's challenge.
  • Approval interval: the verifier-local interval defined in Section 5.1 within which the approval act provably occurred.

2. Design Constraints

C1. In-band. The record arrives inside, or inseparably with, the grant it describes, checkable at request time from retained material, with no token exchange the verifier cannot perform.

C2. Evidence relates to its issuer. The document states, per claim, who issues the evidence and what a verifier can recompute versus must take on attestation. It never implies the evidence outranks its issuer.

C3. Reuse existing vocabulary. acr/amr semantics (OIDC [OpenID.Core]), including the existing amr value "hwk" [RFC8176] and the EAP ACR values [OpenID.EAP] for phishing-resistant authentication, are reused, not reinvented.

C4. Verifier-required, not self-declared. Whether a record is demanded, and at what strength, is the relying party's policy. An asserted value a verifier cannot check is treated as absent.

C5. Parameters, not classes. Timing and artifact binding are parameters of the record rather than new assurance classes; see Section 5.

3. The Approval Record

A grant carries at most one approval record. The grant itself MUST carry a unique issuance identifier (a serial, or issuance timestamp plus id) as a covered field. The REQUIRED form is deliberate. An optional identifier would create a third state, bound but content-only, that every verifier handles differently and no vocabulary names. A grant format that cannot carry an identifier is not excluded; it is priced, through captured binding, under Section 5.

Fields of the record:

The record, attestation included when present, MUST NOT exceed 8 KiB. A verifier that measures a record above the ceiling MUST conclude invalid: the check completed and the record failed it. Unverifiable is reserved for a check that could not be performed.

3.1. Carriage

The record travels beside the grant. Its challenge is the hash of the grant bytes, and a grant already signed by its issuer cannot take a new member without changing the bytes its issuer signed. Keeping the record outside leaves the grant exactly as issued, so any grant format can carry a record, and no grant is re-issued to attach an approval.

In JSON, the grant and its record travel together in a wrapper object with two members, grant and approval_record.

Over HTTP, the record travels in its own header field, Approval-Record, alongside the grant. When the request is signed under the Web Bot Auth protocol (HTTP Message Signatures [RFC9421]), the signature MUST cover the Approval-Record field. A record stripped in transit then breaks the signature, where otherwise it would move the request silently from the record VALID column to record ABSENT (Section 8). The 8 KiB ceiling keeps the field inside common default per-header limits.

4. What the Record Claims, and Does Not

Three claims, three issuers.

Claim 1: This exact issuance was approved. The verifier recomputes the hash of the grant it holds, issuance identifier included, and checks it against the signed challenge. No issuer sits in that path; the claim is recomputable. The choice of semantics is deliberate and stated: the approval binds an issuance, not a content. Two re-issues of identical content are two approvals, and an approval of one is evidence for neither of the other. This is what makes an old record unreplayable onto a re-issued grant. A reader who wants content binding is asking for a different and weaker claim, and this document declines to blur them.

Claim 2: The approval was performed on a hardware authenticator with a user present and verified. The presence and verification flags ride inside the signed authenticator data, checkable against the credential public key. The hardware claim rests on enrollment-time attestation. Its issuer is the authenticator vendor: trust relocated to a party with no stake in any particular grant, not removed. Where the verifier holds no attestation for the credential, the hardware claim is reported as unattested while the presence and verification flags still verify. An authenticator identifier (AAGUID) carried without attestation is self-asserted and proves nothing about the hardware.

Claim 3: That credential belongs to that operator's approver. The residue. The party being graded states it, nothing recomputes it, and an operator could enroll its own credential and manufacture human-looking approvals. A verifier's confidence in claim 3 is exactly its confidence in the operator's enrollment practice, and this document says so plainly.

The honest composite: when all three claims verify, some human holding the enrolled credential approved this exact issuance.

Three partial states are named, so implementations have something to switch on other than a hard fail:

5. Parameters, Not Classes

Two orthogonal parameters appear on every record.

artifact_binding = bound | captured

timing = fresh | standing

A single class value cannot express that bound plus standing is safe while captured plus standing is not. That asymmetry is the argument for parameters.

5.1. Time, and Whose It Is

approved_at has an issuer, and it is the operator. Nothing the authenticator signs carries a clock. authenticatorData holds the RP ID hash, the flags, a signature counter (ordinal rather than temporal, and zero on a great many authenticators) and extension output. clientDataJSON holds the type, the challenge and the origin. So the record's one timestamp rides outside the signed material, and a verifier applying a freshness policy measured from approved_at is trusting the operator's clock while believing it is checking something.

The situation is better than pure trust by exactly one interval. The challenge covers the grant bytes, and the grant carries its issued-at inside those bytes. The assertion cannot exist before the grant it covers, so the covered issued-at is a lower bound on the approval act. The verifier's first sight of the record is an upper bound on its own clock. The approval provably lies inside that interval, and approved_at is an operator-asserted point within it.

Two freshness claims follow, named, so a verifier picks the one it can afford:

  • ASSERTED freshness: a window measured from approved_at. Operator grade: exactly as good as the operator's clock and honesty.
  • CHECKABLE freshness: a window measured from the covered issued-at, evaluated against the verifier's own clock. Recomputable from covered material, with no issuer in the path beyond the grant's own.

Three rules keep the checkable claim checkable:

  1. The upper bound is VERIFIER-LOCAL and MUST NOT be carried as a record field. Two verifiers reading one record compute different intervals, and the narrower one is not more true. Written into the record, the bound becomes one more operator-asserted timestamp under a new name.
  2. First sight means FIRST USE. The interval tightens only for a verifier that keeps state across uses of a standing grant. A one-shot verifier gets the loosest interval, and one-shot is the common case.
  3. The interval is only as independent as the grant's own issuer. Where the operator issues the grant, the covered issued-at is the operator's own assertion and the interval COLLAPSES to operator grade. This caveat sits here, beside the interval, deliberately: a reader who takes the interval from one section and its collapse condition from another will implement the first and not the second.

5.2. Revocation, Measured Against the Same Interval

When the approver credential is revoked, the revocation's effective time is compared with the interval above, not with approved_at:

  • Effective after the verifier's upper bound: the record stands; the credential was valid when the act occurred.
  • Effective before the covered issued-at: invalid; the approver held no valid credential when the act is claimed to have happened.
  • Effective inside the interval: genuinely doubtful, and doubtful verifier-locally in the same way the upper bound is. Rules 1 to 3 above govern it, and the result is the named partial state in Section 4, APPROVED BY AN ENROLLED CREDENTIAL, HOLDER ATTRIBUTION UNRESOLVED AT REVOCATION, which is not a hard fail.

6. Evidence Type: WebAuthn

Challenge construction. challenge = H(canonical grant bytes), where H is SHA-256 [RFC6234], fixed by this document. There is no per-record hash field, because a negotiable hash is a downgrade lever every verifier must defend forever. Agility lives in the registry: a future evidence_type, or a versioned entry, carries a different hash when one is needed.

Verifier procedure.

  1. Recompute H from the grant it holds.
  2. Check equality with clientDataJSON.challenge.
  3. Verify the assertion signature against credential [WebAuthn].
  4. Read the UP and UV flags from authenticatorData.
  5. Evaluate the attestation, from the record or retained from an earlier record for the same credential. With none held, report the hardware claim as unattested.

UV-required versus UP-sufficient maps to acr values per relying-party policy.

8. Worked Example and Failure Matrix

One agent. Operator anchor on one side (a JWK Set at a Signature-Agent URI the operator controls), approval-bound grant on the other. Materials are real WebAuthn assertions over real grant hashes, produced by a working prototype and labeled as such. The prototype mints a JCS grant, runs the passkey ceremony over its hash and verifies per Sections 4 to 6. The failure set is 13 vectors, one fault each, with each basis quoting the section it tests. The current fixture is a 1Password passkey with attestation "none", so the verifier reports claim 2's hardware claim as unattested.

Anchor outcomes are three: VALID; LAPSED, a conclusive outcome where the anchor expired and the verifier knows it; and UNREACHABLE, which is not an outcome because the check never completed. Anchor changed-hands is deliberately absent from the matrix. It is an ordering question between two events, which is record-layer territory, referenced rather than absorbed.

Cells report per-claim outcomes, not a single verdict.

Table 1
Record VALID Record ABSENT Record INVALID
Anchor VALID Claim 1 ok, Claim 2 ok, Claim 3 ok: accountable operator, human-approved this exact issuance Accountable operator; approval strength unknown (status quo) Invalid record. Forgery is one possible cause. The one cell treated worse than absent.
Anchor LAPSED (conclusive) Claim 1 ok, Claim 2 ok, Claim 3 rests on an expired anchor: APPROVED BY AN ENROLLED CREDENTIAL, HOLDER ATTRIBUTION EXPIRED. A named partial state, not a hard fail. Operator accountability expired; nothing else asserted Invalid record, expired anchor: nothing, loudly
Anchor UNREACHABLE (no outcome) Claim 1 ok, Claim 2 ok, Claim 3 unevaluated: APPROVED BY AN ENROLLED CREDENTIAL, HOLDER UNATTRIBUTED. A named partial state, not a hard fail. NO VERDICT: the check never completed Invalid is invalid regardless of anchor state

Non-independence warning. The two axes share a trust root. Claim 3's issuer is the anchor party, so anchor-valid is a precondition of claim 3, not corroborating evidence for it. The anchor column and the claim 3 row are the same trust, spent twice; a reader must not sum them.

9. Implementation Status

Per [RFC7942], to be removed before publication as an RFC. One prototype exists: a Cloudflare Worker that mints a JCS grant, performs the WebAuthn ceremony over its hash, verifies per Sections 4 to 6, and generates the 13-vector failure set from a real ceremony. Coverage: one 1Password passkey fixture, attestation "none". The vectors are available from the authors on request.

10. Security Considerations

11. IANA Considerations

12. Privacy Considerations

The record exposes an approval event and a credential identifier to every verifier of the grant. Three facts follow.

The credential public key is the sharpest exposure. A verifier needs it to check the assertion, so the record cannot withhold it. Where the approver is the person the agent acts for, every verifier receives the same key and can link all of that person's agents through it. The raw key is exactly the correlator the at-rest rule below exists for.

Pairwise credentials per relying party are INCOMPATIBLE with standing grants. They require enumerating the verifiers at approval time, and the reason standing grants exist is that you cannot.

The workable lever is at rest, not on the wire. A verifier SHOULD store H(credential_id, per-verifier salt) rather than the raw identifier. Per-verifier continuity survives, cross-verifier correlation ends, and C1 is untouched because nothing about the record changes. The protocol cannot enforce this, since it is storage policy, which is why it is a verifier SHOULD rather than a wire mechanism.

13. References

13.1. Normative References

[OpenID.Core]
OpenID Foundation, "OpenID Connect Core 1.0", , <https://openid.net/specs/openid-connect-core-1_0.html>.
[OpenID.EAP]
Jones, M.B., "OpenID Connect Extended Authentication Profile (EAP) ACR Values 1.0", , <https://openid.net/specs/openid-connect-eap-acr-values-1_0.html>.
[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <https://www.rfc-editor.org/info/rfc2119>.
[RFC6234]
Eastlake 3rd, D. and T. Hansen, "US Secure Hash Algorithms (SHA and SHA-based HMAC and HKDF)", RFC 6234, DOI 10.17487/RFC6234, , <https://www.rfc-editor.org/info/rfc6234>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <https://www.rfc-editor.org/info/rfc8174>.
[RFC8176]
Jones, M., Hunt, P., and A. Nadalin, "Authentication Method Reference Values", RFC 8176, DOI 10.17487/RFC8176, , <https://www.rfc-editor.org/info/rfc8176>.
[RFC8785]
Rundgren, A., Jordan, B., and S. Erdtman, "JSON Canonicalization Scheme (JCS)", RFC 8785, DOI 10.17487/RFC8785, , <https://www.rfc-editor.org/info/rfc8785>.
[RFC9421]
Backman, A., Ed., Richer, J., Ed., and M. Sporny, "HTTP Message Signatures", RFC 9421, DOI 10.17487/RFC9421, , <https://www.rfc-editor.org/info/rfc9421>.
[WebAuthn]
W3C, "Web Authentication: An API for accessing Public Key Credentials Level 3", W3C Web Authentication Level 3, <https://www.w3.org/TR/webauthn-3/>.

13.2. Informative References

[AAuth]
Hardt, D., "AAuth Protocol", Work in Progress, Internet-Draft, draft-hardt-oauth-aauth-protocol-11, , <https://datatracker.ietf.org/doc/html/draft-hardt-oauth-aauth-protocol-11>.
[I-D.brezun-human-continuity-http]
Brezun, A., "Human Continuity for HTTP", Work in Progress, Internet-Draft, draft-brezun-human-continuity-http-00, , <https://datatracker.ietf.org/doc/html/draft-brezun-human-continuity-http-00>.
[I-D.ietf-webbotauth-httpsig-protocol]
Meunier, T. and S. Major, "HTTP Message Signatures for automated traffic", Work in Progress, Internet-Draft, draft-ietf-webbotauth-httpsig-protocol-01, , <https://datatracker.ietf.org/doc/html/draft-ietf-webbotauth-httpsig-protocol-01>.
[I-D.morrison-acts-for]
Morrison, B., "Acts-For: Signing In to a Website with a ~handle, for a Person or an Agent Acting for One", Work in Progress, Internet-Draft, draft-morrison-acts-for-01, , <https://datatracker.ietf.org/doc/html/draft-morrison-acts-for-01>.
[IPSIE]
IETF, "IPSIE working group documents", <https://datatracker.ietf.org/wg/ipsie/documents/>.
[RFC7942]
Sheffer, Y. and A. Farrel, "Improving Awareness of Running Code: The Implementation Status Section", BCP 205, RFC 7942, DOI 10.17487/RFC7942, , <https://www.rfc-editor.org/info/rfc7942>.

Appendix A. Open Issues

  1. Intended status: Standards Track or Experimental.
  2. Whether the two new registries should merge into one.
  3. A second, attested fixture (a security key with packed attestation) for the worked example.
  4. Final names for the approval_record member and the Approval-Record field.

Acknowledgments

Drew, Alter Meridian Pty Ltd. Aurelien Brezun: parameters, not classes; the unique-human boundary.

Authors' Addresses

Joshua Ashcroft
Kipple Labs
Blake Morrison
Alter Meridian Pty Ltd
Nick Mathews
Agentic Verification Architecture