Review · November 21, 2026
An Audit Record Format for AI Agent Authorization Decisions
What problem is this solving?
AIMS says deployments MUST produce tamper-evident audit logs with seven minimum fields. It deliberately does not say what those logs look like. This draft supplies the format.
The unique part is not the seven fields. The draft puts the authorization decision and the observed effect in the same signed record. A "deny" that was followed by a write shows up as a disagreement in signed bytes. A log that only records the decision would show a clean denial.
How does it fit in?
It defines one new artifact and no protocol. There are no new endpoints, tokens or flows, and no IANA actions.
It depends on zero OAuth RFCs. The normative dependencies are AIMS, the WIMSE identifier draft, JCS (RFC 8785), I-JSON (RFC 7493) and RFC 3339. The envelope and statement formats come from outside the IETF: in-toto Statements signed in a DSSE envelope.
It introduces one new role, the observer. This is a party watching from a layer the agent cannot address, such as a hypervisor or kernel-level supervisor. AIMS has no such role. The decision point's own report is treated as an unverified claim next to what the observer saw.
Charter fit in WIMSE is weak. AIMS explicitly declines to standardize the record format. Datatracker now shows this draft as replaced by draft-gilda-audit-agent-audit-record, so the author has already moved it toward the proposed AUDIT working group. That is the better home. Note that the AUDIT charter puts "how and when to create the records", and by which entity, out of scope. This draft's strongest tier depends entirely on who created the record and from where.
What does the record contain?
Each record covers one request and one interval of observed state. It has seventeen required members and one optional one. The main groups:
- Who and what: agent identifier, credential digest, authentication method, delegated subject, resource and a digest of the call arguments.
- Decision: permit, deny or permit-with-conditions, plus whether the decision point actually evaluated the request or failed open or closed.
- Effect: a before-state root, an after-state root, and an ordered list of file writes with pre and post digests.
- Agreement: derived from decision and effect. A deny with a write is
disagree. A permit with no write isnot-exercised. A write the observer cannot attribute isindeterminate. - Posture: the posture the decision point read next to the posture the observer saw.
- Remediation: revocation signals with the time received and the time enforced.
A verifier checks the signature, then recomputes the request digest, the write chain and both agreement values from other members of the record. An inconsistent record is rejected even when the signature is valid.
Records are sorted into two tiers. authoritative requires five things: an observer below the agent, a commitment to the before-state signed before the interval opened, a named path scope, complete coverage and digest-bound arguments. Everything else is voluntary. The verifier recomputes the tier and never trusts the declared value.
The draft is honest about the limits. A producer holding two keys can satisfy every clause. Nothing detects a record that was never written.
Significant interoperability concerns
The effect model is a filesystem. Effects are writes to paths, with state roots before and after. Most agent actions are API calls to Google, Salesforce or Slack, or MCP tool calls. No observer can compute a before-state root for a remote SaaS API. Those records can never be authoritative, and their effect can only be "none" or "unknown". The format fits a sandboxed coding agent and little else.
The recomputes are not specified enough to implement. The draft says the write chain "MUST reproduce the after-state root from the before-state root". It never defines how a root is computed from a set of writes. The "empty-tree constant" is not defined. Neither are the bytes that credentialDigest, policyDigest and authorityDigest are taken over, or the "explicit deny-all document". Two implementations will produce different roots for the same filesystem.
The prior commitment signature is undefined. priorCommitment carries a keyid and a sig. The draft does not say what bytes are signed or with what algorithm. Key discovery for the observer is out of band.
The request digest preimage does not match the record. It is defined over resourceId and resourceKind. The record has resource.id and resource.kind. The exact JSON object to hash needs to be written out.
The example contradicts the text. Section 4.3 says the correlation identifier is the first subject entry's name. The examples use urn:example:corr:7f3a for the subject and req:7f3a91c4 for correlation.id.
Hash agility is half done. One hashAlgorithm governs every digest, but the subject digests are keyed sha256 and the text refers to "its sha256 digest" throughout.
The identifiers and vocabularies live in the author's GitHub. The predicate type URI is under probityai.github.io. The closed vocabularies come from a GitHub repository cited as an informative reference, yet a verifier MUST reject values outside them. The conformance vectors are described in a table and not yet published. The URI is not frozen.
The posture values are not AIMS posture. no_network, allowlist, sinkhole and unsafe_bypass_egress describe sandbox network egress. AIMS posture covers much more. Any deployment with a different posture model cannot fill in this required field.
Records are not stable across revisions. Three revisions shipped in two weeks. A valid -01 record is malformed under -02, with the same predicate type URI.
Mechanisms that could be removed
- The declared
tier. The verifier MUST recompute it and MUST NOT trust it. A declared value that fails is malformed. The member carries no information. - The declared
agreementvalues. Same argument. They are pure functions of other members. one-sided. A reserved vocabulary value that makes any record carrying it malformed.oversight. Optional, producer-asserted, never resolved by the verifier, and built on a -00 individual terminology draft.doesNotAssert. A required array of free-form strings. No verifier can act on it.externalAnchor. No validation rule is defined. The draft's own test vector lists the result as indeterminate.
A smaller draft would keep the canonicalization rules, the seven-field mapping, and the decision and effect pair.
Writing quality
The writing is precise and well above average for an individual draft. BCP 14 keywords are used carefully. Almost every normative rule points to a test vector, and each reject vector is a single mutation of an accept vector. That is a good practice more drafts should copy.
The prose is stylized. Sentences like "A disagreement is the record working" recur, and several sections restate earlier rules in different words. Section 4.6 and Section 5.8 are mostly repetition. The document is 31 pages and could lose a quarter of that.
Some claims run ahead of the mechanism. Section 6.7 says the recomputes address "a producer that signed a false record". They catch an inconsistent record. A producer that lies consistently passes every check, as Section 6.2 admits.
Who is the author?
Sankalp Gilda is the sole author, with no affiliation listed. Datatracker shows zero RFCs and no expired drafts. There are two active drafts, both renamed for the proposed AUDIT group: draft-gilda-audit-agent-audit-record, which replaces this one, and draft-gilda-audit-restoration-evidence. This is a new IETF author who is writing actively and revising fast.
So what should the working group do with this?
Take the central idea seriously. Recording the decision without the effect produces a clean audit trail for an enforcement failure. AIMS field four does not address that, and it should.
The format itself is not ready. It needs a defined state-root construction, an effect model that covers remote API calls, and identifiers that do not live in one person's GitHub namespace. Review effort should go to the replacement draft in AUDIT, not to this revision.