| Internet-Draft | AI Governance Verified | July 2026 |
| Hillier | Expires 25 January 2027 | [Page] |
This document specifies a verification standard for the cryptographic attestation of agentic AI governance in regulated industries. It defines the Verification Reconciliation Object (VRO), the issuing-partner framework, the eight control areas through which AI governance posture is reconciled, three maturity-attestation levels (Documented, Operational, Adversarial-ready), and the cryptographic continuity requirements that together produce deterministic, independently reconstructable, auditor-grade attestations of agentic AI governance. The standard sits beneath ISO/IEC 42001:2023, the NIST AI Risk Management Framework, and other agentic AI governance frameworks, and produces the verifiable artefact those frameworks were designed to imply but do not deliver.¶
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 25 January 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Agentic artificial intelligence is now operative across the workplace at a scale that exceeds the control envelope of every previously published governance framework. ISO/IEC 42001:2023 [ISO42001] specifies a management system for AI but does not produce a verifiable artefact. The NIST AI Risk Management Framework [NIST-AI-RMF] provides a functional taxonomy but issues no certification or attestation. The European Union AI Act [EU-AI-ACT] establishes obligations and prohibitions but leaves verification of compliance to national competent authorities and self-attestation. No published standard issues a cryptographically anchored, deterministically reproducible monthly artefact of agentic AI governance.¶
This document closes that gap. It defines the verification artefact, the issuing-partner framework, the evidence requirements across eight control areas, the maturity-level attestation methodology, and the cryptographic continuity requirements that together produce a deterministic, auditor-grade AI governance attestation.¶
The urgency of that artefact has been underscored by a class of failure now observed in practice: an autonomous system reaching an assigned objective through a consequence its operators neither authorised nor observed in time — including, in a controlled capability evaluation, an agent escaping its intended execution boundary and acting on external infrastructure without human authorisation, detected only after the fact. Governance that is documented but never adversarially exercised does not detect this class. The Adversarial-ready Maturity Level (Section 5) and the evaluation-time containment requirement (Section 4.9) introduced in this document address it directly.¶
This document does not replace ISO/IEC 42001, the NIST AI Risk Management Framework, the EU AI Act, or any national framework. It sits beneath them and produces the artefact each was designed to imply but does not deliver. Where this document and any normative framework cited herein conflict on operational content, the cited framework prevails.¶
This document applies to organisations that operate, integrate, deploy, or expose AI systems — whether developed internally, procured from third-party model providers, or consumed via API. It does not specify AI model architecture, training procedure, or evaluation methodology. It specifies the verification of governance applied to AI use, not the AI itself.¶
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.¶
For the purposes of this document, the following definitions apply.¶
The organisation whose AI governance posture is the subject of verification.¶
Any deployed system, application, or service whose behaviour incorporates the use of one or more machine-learning models, irrespective of model architecture or training methodology.¶
An AI System that takes actions in the operating environment, generates outputs that influence downstream decisions, or operates with reduced or absent human-in-the-loop supervision.¶
An external party supplying a foundation model, fine-tuned model, or model-as-a-service to the Subject Entity.¶
The set of integrations, data sources, user populations, regulatory obligations, and risk attributes within which an AI System is operated.¶
The composite of policies, controls, telemetry, evidence sources, and review cadences through which the Subject Entity governs the use of its AI Systems.¶
The deterministic, cryptographically anchored output of a conforming AI governance attestation under this standard.¶
A counterparty designated by the protocol operator to act as a co-issuer of VROs under this standard within a defined market or scope.¶
The contiguous time interval over which a VRO asserts conformance.¶
The cryptographic operation that binds a VRO to an immutable public settlement layer at issuance and at supersession.¶
One of three levels (Documented, Operational, Adversarial-ready) defined in Section 5.¶
The lifecycle event by which a new VRO replaces a prior VRO.¶
Conforming attestations under this standard are produced by a verification infrastructure organised as five architectural components. Internal design, scoring methodology, and calibration logic are not in scope for this document.¶
Conforming attestations are deterministic. Given the same Conformance Claims and the same Evidence Artefacts processed through the same Attestation Protocol version, the same VRO SHALL be produced. Determinism applies to the verification operation, not to the AI systems being verified.¶
Sections 4.1 to 4.8 specify the eight control areas through which agentic AI governance is reconciled under this standard.¶
The Subject Entity maintains a current inventory of AI Systems in operation across its workforce, integrations, and infrastructure, and detects use of unsanctioned or undisclosed AI Systems within its operating environment.¶
AI System register; identity-provider telemetry of AI service authentications; endpoint or network telemetry of model API egress; sanctioned-application register; shadow-AI detection output; periodic reconciliation reports.¶
Reconciliation of declared inventory against detected use across the Attestation Period; exception cases reconciled against the disclosure or remediation register.¶
Anchored at Attestation Period start and end. Material expansions of inventory require a supersession anchor.¶
The Subject Entity classifies each AI System by use-case category and risk tier, and records the classification together with the rationale and the residual-risk position.¶
Use-case classification register; risk assessment artefact for each AI System; risk-tier policy artefact; review records evidencing periodic reassessment; exception register for ungoverned use cases.¶
Reconciliation of declared classifications against the policy taxonomy; reconciliation of risk tier against deployment context evidence; identification of classification drift over the Attestation Period.¶
Anchored at issuance; supersession on material change in deployment context, risk tier, or use-case scope.¶
The Subject Entity records and maintains provenance evidence for each AI System in use, including model identity, model version, Model Provider identity, training data disclosures (where available), and update or fine-tuning lineage.¶
Model registry entries with provider, version, and lineage data; Model Provider transparency reports or model cards where supplied; training data attestations where available; fine-tuning records; vendor change log.¶
Reconciliation of recorded provenance against AI System operational state; reconciliation of fine-tuning lineage against change-management evidence; identification of model-version drift across the Attestation Period.¶
Anchored at issuance and at each material model-version change or Model Provider change.¶
The Subject Entity restricts use of AI Systems to those that have been explicitly sanctioned for the applicable user population and use-case context, consistent with the asserted Maturity Level.¶
Sanctioned-application allow-list policy; endpoint or network enforcement evidence; exception register; user-population scope evidence; periodic review records.¶
Reconciliation of declared allow-list against enforcement telemetry; reconciliation of exception cases against the exception register; identification of unsanctioned use over the Attestation Period.¶
Anchored at issuance and on material allow-list changes.¶
The Subject Entity governs the content of prompts submitted to AI Systems and outputs produced by AI Systems, including controls preventing disclosure of sensitive data to external AI Systems and controls preventing high-risk output content from entering downstream processes.¶
Prompt-content policy artefact; output-content policy artefact; prompt-monitoring telemetry; output-review telemetry; data-loss-prevention rules applied to AI traffic; high-risk content exception register.¶
Reconciliation of declared content controls against monitoring telemetry; reconciliation of exception handling against review evidence; identification of control bypass or drift over the Attestation Period.¶
Anchored at issuance and on each material change in control scope, sensitive-content taxonomy, or enforcement state.¶
The Subject Entity enforces identity and access controls on the use of AI Systems consistent with the asserted Maturity Level, including authentication, authorisation, multi-factor enforcement, and segregation between human and machine principals.¶
Identity-provider telemetry for AI service authentications; access assignment register for AI Systems; multi-factor enrolment coverage report; service-account inventory; access review records.¶
Reconciliation of declared access model against assignment register; reconciliation of multi-factor enforcement against identity-provider evidence; reconciliation of service-account use against the inventory and policy.¶
Anchored at issuance and at the conclusion of each scheduled access review cycle.¶
The Subject Entity captures, retains, and protects logs of AI System use sufficient to permit retrospective reconstruction of governance-relevant events, with retention and integrity properties consistent with the asserted Maturity Level.¶
Logging policy artefact; log content schema; log retention configuration; log-integrity attestation; access-control evidence for log stores; sampling or audit records.¶
Reconciliation of declared logging scope against captured content; reconciliation of asserted retention against log-store configuration; reconciliation of asserted integrity properties against evidence of log-store immutability or chain protection.¶
Anchored at issuance and on material changes to logging scope, retention, or integrity configuration.¶
The Subject Entity operates an incident-response capability for AI-related events, including hallucination, output failure, prompt-injection, data exfiltration, model drift, and high-impact misuse, with defined escalation paths and post-event review consistent with the asserted Maturity Level.¶
AI incident-response policy; incident register; incident classification taxonomy; escalation records; post-event review reports; remediation evidence; drift-monitoring telemetry.¶
Reconciliation of declared response capability against the incident register over the Attestation Period; reconciliation of escalation evidence against the declared escalation paths; reconciliation of remediation evidence against committed actions.¶
Anchored at issuance and at the conclusion of each material incident response or post-event review cycle.¶
Where a Subject Entity conducts capability evaluations, red-team exercises, or other adversarial tests of an AI System --- particularly tests that deliberately reduce guardrails or grant elevated capability to the system under test --- the containment of the evaluation environment is itself a governance-relevant control and SHALL be within the scope of this control area.¶
At Maturity Level Adversarial-ready, the Subject Entity SHALL verify, in real time during the evaluation, that the system under test remains within its intended execution boundary: that network egress is restricted to the declared allow-list, that no capability escalation beyond the evaluation's declared scope occurs, and that any action reaching external infrastructure is reconciled against the evaluation's authorised scope before it takes effect. Detection of boundary escape SHALL be treated as a reportable incident under this control area, with the containment failure --- not only the model behaviour --- recorded in the incident register.¶
Evidence Categories for this requirement include: evaluation-environment containment policy; egress-control configuration and telemetry for the evaluation environment; real-time boundary-monitoring records; and post-evaluation reconciliation of intended versus actual system conduct.¶
At Maturity Levels Documented and Operational, the Subject Entity SHALL record the containment posture of its evaluation environments; real-time boundary reconciliation is RECOMMENDED but not required below Adversarial-ready.¶
A conforming VRO under this standard SHALL attest a Maturity Level for each of the eight control areas. Different control areas MAY attest at different Maturity Levels within a single VRO; the overall VRO attestation is the minimum Maturity Level attested across the eight control areas unless otherwise asserted.¶
Reflects the baseline expectation that governance content exists, that policy artefacts are written, that an inventory is maintained, and that a designated owner is accountable. Evidence requirements at this level emphasise the existence of declared content and basic operational artefacts over an Attestation Period of at least three (3) consecutive months.¶
Reflects the expectation that controls are not only documented but exercised: telemetry collected, periodic reviews occurring, exceptions recorded and handled, and the governance surface responding to material change. Evidence requirements add depth of telemetry, review-cadence evidence, exception handling, and continuity over an Attestation Period of at least six (6) consecutive months.¶
Reflects the expectation that governance withstands adversarial conditions: prompt-injection attempts, model-drift events, data-exfiltration attempts via AI channels, sophisticated misuse, and dependency failures at the Model Provider. Evidence requirements add continuity, defence-in-depth evidence, red-team or adversarial evaluation attestation where in scope, and continuous reconciliation over an Attestation Period of at least twelve (12) consecutive months.¶
A Subject Entity that progresses to a higher Maturity Level for any control area SHALL be issued a superseding VRO recording the progression. The prior VRO is preserved and marked as superseded.¶
A conforming VRO under this standard SHALL contain, at minimum:¶
A VRO MAY be revoked by the Issuing Partner upon determination of material non-conformance, evidence falsification, undisclosed incidents, or other circumstances rendering the original attestation unreliable. Revocation does not delete the VRO; it records a revocation state, the revocation reason class, and the Anchor Event binding the revocation to the public settlement layer.¶
Each issued VRO SHALL be registered in the public attestation registry.¶
An organisation seeking designation as an Issuing Partner under this standard SHALL demonstrate, at minimum:¶
An Issuing Partner SHALL NOT, for a given Subject Entity engagement, simultaneously act as the implementing vendor, deployment integrator, or operator of the AI Systems being verified.¶
Each VRO SHALL be cryptographically anchored to an immutable public settlement layer at the Anchor Event. The hash committed at the Anchor Event SHALL be a one-way function of the VRO content, Issuing Partner identity, and timestamp, computed under a digest algorithm of at least 256-bit strength.¶
A VRO issued under this standard SHALL remain a conforming artefact across regulatory regime changes occurring within or after the Attestation Period.¶
The Anchor Event binding SHALL remain independently verifiable in the event of a Model Provider ceasing to operate, withdrawing a model, or being acquired or restructured. VROs issued during the operating life of a withdrawn model are not retroactively invalidated.¶
This standard is interoperable with adjacent frameworks. Conforming VROs MAY be referenced within audit, certification, and regulatory artefacts produced under:¶
This revision maps the following obligations explicitly. EU AI Act [EU-AI-ACT] Article 14 (human oversight) is evidenced through Areas 6 and 8 and the evaluation-time containment requirement (Section 4.9); Article 50 (transparency obligations) is evidenced through Areas 3 and 7. NIST AI Risk Management Framework [NIST-AI-RMF] Manage and Measure functions map to the incident, drift, logging, and containment evidence categories. Reconciliation Outputs and VRO Anchor Events MAY be notarised as transparent statements under the SCITT architecture [I-D.ietf-scitt-architecture], and an agent-action reconciliation performed under the Attestation Reconciliation Protocol [I-D.hillier-scitt-arp] MAY supply the real-time containment evidence required by the evaluation-time containment requirement.¶
An attestation artefact MAY claim conformance to this standard if and only if it satisfies every requirement specified in Sections 3 through 8. Partial conformance is not recognised. Variant conformance to a subset of control areas without the full eight-area scope is not recognised.¶
The public attestation registry constitutes the authoritative record of issued VROs.¶
This document has no IANA actions.¶
Agentic AI governance operates under adversarial conditions distinct from traditional cybersecurity. Implementations of this standard SHOULD pay particular attention to prompt-injection resistance, model-drift detection, and exfiltration paths through AI channels that may bypass traditional data-loss-prevention controls.¶
Issuing Partners are required by Section 7 to be independent from the Subject Entity and from Model Providers whose models are within attestation scope.¶
The Anchor Event binding SHOULD use a digest algorithm of at least 256-bit strength and a public settlement layer with no single private operator capable of extinguishing the binding.¶
This standard does not address the correctness or safety of the AI Systems being governed. It addresses the verifiability of governance applied to those systems.¶
This appendix is informative.¶
The requirements of this document, and in particular the Adversarial-ready Maturity Level and the evaluation-time containment requirement of Section 4.9, are motivated by a class of failure in which an autonomous system reaches an assigned objective through a consequence that its operators neither authorised nor observed in time.¶
A representative instance of the class, as disclosed publicly, is a cyber-capability evaluation in which an autonomous system operating under reduced guardrails escaped the evaluation's execution boundary by exploiting an unremediated vulnerability in a supporting service, obtained network egress its containment had assumed impossible, and achieved code execution on external infrastructure, retrieving material from outside its authorised scope --- without human authorisation, and detected only after the fact.¶
The generalisable properties of the class are: that the execution boundary was assumed rather than continuously proven; that the authority under which the system acted was not scoped and bound to its actual conduct; and that there was no independent, real-time reconciliation of the system's claimed activity against its actual activity while the action could still be refused. This document addresses the first and third through the evaluation-time containment requirement and the reconciliation-based evidence model; the second is addressed at the protocol layer by the Attestation Reconciliation Protocol [I-D.hillier-scitt-arp].¶
RFC Editor: please remove this section before publication.¶