compliance · NIST AI RMF
Wytness and NIST AI RMF
Subcategory evidence for the AI agents inside your risk management programme.
NIST AI RMF 1.0 is a voluntary framework: four functions — GOVERN, MAP, MEASURE, MANAGE — with 72 subcategories your organisation applies to its own AI risk management. Where the systems in scope are AI agents, Wytness evidences the subcategories that rest on a record of what the agents did — the inventory, roles, intended use and declared scope, production monitoring, the integrity of the record, data classes, deployment approval and third-party model usage. The risk management itself — risk tolerance, impact assessment, testing and evaluation, risk treatment and incident response — stays yours. The other 61 subcategories are listed below with the reason the pack does not evidence them.
Sources: Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (January 2023, PDF) · NIST AI RMF programme page
the artefact
What's in the pack
A NIST AI RMF pack contains the agent and tool inventories, monitoring evidence per subcategory, signed event samples, the LLM provider and model mix, and a subcategory-by-subcategory narrative you set against your own RMF profile, and the list of subcategories it does not evidence, with the reason for each.
Every claim in the pack is traceable: each control cites the events behind it, and each event carries the signature your own key produced. The pack is sealed, so an auditor can confirm it reached them unaltered.
A real sealed pack is published — nothing here is a mock-up. Download the specimen and check it yourself on the verify page, in your browser, without uploading it. It is a SOC 2 (Trust Services Criteria) pack — the seal and the per-event signatures work identically whichever framework you generate.
// What the pack asserts about itself:
// control_mapping[] — 11 criteria, each carrying
// its evidence and a metric measured over your period:
…// + 7 more, in the table below
// Every control cites events an auditor can re-verify:
the mapping
What we map to what
| Subcategory | Name | What Wytness provides |
|---|---|---|
| GOVERN 1.6 | AI system inventory | The Agent Registry inventories every agent seen in the event stream — an agent that emits events is added to the register as auto_discovered so none runs unrecorded; each row carries an asset type (agent, generative-AI application or ML model), version label, lineage (derived_from), risk classification, approval state and its declared tools and permissions. The inventory exports as JSON or CSV and as a signed register PDF. |
| GOVERN 2.1 | Roles, responsibilities and lines of communication | Each agent in the Agent Registry carries an assigned operator (a named user of the organisation) and an approval record (approved_by, approved_at); every event records human_operator_id; four-tier role-based access control (viewer / member / admin / owner) is enforced server-side on every route; each register change is appended to the platform audit log with actor, target, and the fields changed. |
| MAP 1.1 | Intended purposes and deployment context | The Agent Registry documents each agent's intended use and description beside its risk classification (minimal / limited / high / prohibited, with a justification required for high and prohibited), its EU AI Act category, and its approval state; every event carries agent_id, agent_version and agent_instance_id, so the customer can set recorded use against documented intended use. |
| MAP 3.3 | Targeted application scope | The Agent Registry holds each agent's declared tools and declared permissions beside its intended use; Preflight compares the recorded tape against the declared set, and the capability_profile_deviation anomaly rule fires when an agent's recent observed actions fall outside its seven-day baseline. The scope is the customer's declaration; Wytness records it and reports departures from it. |
| MEASURE 2.4 | Production monitoring of functionality and behaviour | 7 anomaly-detection rule families run against the live event stream (first-time tool use, capability drift, frequency spikes, error spikes, restricted-data access, off-hours activity, hash-chain breaks) with automated severity-classified alerting; the dashboard surfaces event rate, error rate, anomaly count, and chain-integrity status with an explicit as_of freshness signal, with per-agent and per-session drilldown. |
| MEASURE 2.7 | Security and resilience | Wytness evidences the security and resilience of the record, not of the model: the hash chain is verified link by link per agent (live and on demand via /verify), per-event Ed25519 signatures are checked against the customer's registered public keys, and a three-way ingest / query-store / archive reconciliation runs on a schedule; the pack states each control's pass / fail / no-data status for the period. Adversarial testing and robustness or safety evaluation of the model remain the customer's and are not performed by Wytness. |
| MEASURE 2.8 | Transparency and accountability | Every tool call, resource access, and inter-agent message is logged and hash-chained (tamper-evident) with agent_id, agent_version and human_operator_id, and carries a per-event Ed25519 signature, so each action is attributable to an agent and an accountable operator; chain breaks surface as anomalies and a separate platform audit log records administrator actions. Evidence Packs document the period for auditors — agent and tool inventories, per-control evidence with a measured metric, anomaly history, signed event samples and the LLM provider and model mix — paired with a sealed JSON an auditor verifies offline without a Wytness account. Events are retained for seven years by a storage lifecycle policy on Wytness-hosted storage, or under the customer's own policy on bring-your-own-storage; alteration is made evident rather than prevented. |
| MEASURE 2.10 | Privacy risk | Every event records inputs_source and inputs_classification — the provenance and sensitivity class of the data the agent acted on, not the raw payload — and the Agent Registry tags each agent with the data classes it accesses (data_classes_accessed); an event classified restricted raises a restricted_data anomaly. Where the customer enables PII keys, personal data in payloads is replaced with HMAC pseudonyms in the customer's runtime before egress and the token map is encrypted to the customer's X25519 public key, so Wytness stores pseudonyms it cannot reverse. The privacy impact assessment itself is the customer's. |
| MANAGE 1.1 | Determination to proceed with deployment | The Agent Registry records an approval state for every agent (discovered → intake → development → validation → approval → deployment → retirement) with approved_by and approved_at; agents seen in the event stream are added to the register as auto_discovered so none runs unrecorded, and each approval change is logged separately as agent_approval_changed with its from / to states. Wytness records the approval; it does not block an unapproved agent. |
| MANAGE 3.1 | Third-party resources monitored | Wytness records which third-party model providers and models each agent called — the pack's LLM-usage section reports calls, tokens and cost per model and the per-provider call mix for the period — and inventories every tool an agent invokes, raising a new_tool anomaly on first use. Supplier due diligence, contracts and the risk controls applied to third parties remain the customer's. |
| MANAGE 4.1 | Post-deployment monitoring | Post-deployment monitoring runs on the live event stream: the anomaly rule families alert with a severity, each alert carries a resolution state (pending / acknowledged / resolved) backed by anomaly_resolutions, and connectors push alerts to the customer's own SIEM, ITSM or chat (Sentinel, Splunk, ServiceNow, Slack, generic webhook). Appeal and override, decommissioning, incident response, recovery and change-management plans are the customer's; Wytness evidences the monitoring half. |
the boundary
What Wytness does not cover
| Subcategory | Name | Why the pack does not evidence it |
|---|---|---|
| GOVERN 1.1 | Legal and regulatory requirements understood and documented | A legal register is the customer's document. |
| GOVERN 1.2 | Trustworthy-AI characteristics in organisational policy | Policy content; Wytness records agent activity, it does not hold the customer's policies. Shield policies are technical rules in observe mode. |
| GOVERN 1.3 | Level of risk management set by risk tolerance | Risk tolerance is the customer's determination. |
| GOVERN 1.4 | Risk management process established through transparent policies | Same as GOVERN 1.2. |
| GOVERN 1.5 | Ongoing monitoring and periodic review of the risk process | Review of the process is a governance activity; monitoring of the agents is evidenced under MEASURE 2.4. |
| GOVERN 1.7 | Decommissioning and phase-out processes | The register records a `deprecated` state (MANAGE 1.1 row); the decommissioning process is the customer's. |
| GOVERN 2.2 | AI risk management training | Training records are outside the product. |
| GOVERN 2.3 | Executive responsibility for AI risk decisions | A leadership accountability statement; no Wytness mechanism. |
| GOVERN 3.1 | Diverse team informs risk decisions | Team composition is outside the product. |
| GOVERN 3.2 | Roles for human-AI configurations and oversight defined by policy | The policy defining oversight roles is the customer's; the assigned operator per agent is evidenced under GOVERN 2.1. |
| GOVERN 4.1 | Critical-thinking and safety-first culture | Culture; no Wytness mechanism. |
| GOVERN 4.2 | Teams document and communicate risks and impacts | Impact documentation is the customer's. |
| GOVERN 4.3 | Practices for testing, incident identification and information sharing | Connectors push anomalies to the customer's own tools (MANAGE 4.1 row); the organisational practice is the customer's. |
| GOVERN 5.1 | External feedback collected and integrated | A feedback channel; Wytness has none. |
| GOVERN 5.2 | Adjudicated feedback incorporated into design | Same as GOVERN 5.1. |
| GOVERN 6.1 | Policies for third-party software and data risk | Policy content; third-party usage is evidenced under MANAGE 3.1. |
| GOVERN 6.2 | Contingency processes for third-party failures | Contingency planning is the customer's. |
| MAP 1.2 | Interdisciplinary actors and competencies | Team composition is outside the product. |
| MAP 1.3 | Organisational mission and goals for AI | Strategy document; no Wytness mechanism. |
| MAP 1.4 | Business value of the AI use | Business case is the customer's. |
| MAP 1.5 | Risk tolerances determined | Risk tolerance is the customer's determination. |
| MAP 1.6 | System requirements elicited | Requirements documents are the customer's. |
| MAP 2.1 | Tasks and methods defined | The customer defines tasks and methods; the register's asset type records the kind of system, evidenced under GOVERN 1.6. |
| MAP 2.2 | Knowledge limits and human oversight of output documented | A design document; no Wytness mechanism. |
| MAP 2.3 | Scientific integrity and TEVV considerations | Wytness does not test or evaluate the model. |
| MAP 3.1 | Potential benefits examined | Assessment is the customer's. |
| MAP 3.2 | Potential costs examined | Same as MAP 3.1. |
| MAP 3.4 | Operator proficiency processes | Competence records are outside the product. |
| MAP 3.5 | Human oversight processes defined | The oversight process definition is the customer's; Wytness records who operated each agent (GOVERN 2.1 row). |
| MAP 4.1 | Legal risks of components mapped | Legal risk mapping is the customer's. |
| MAP 4.2 | Internal risk controls for components identified | Risk control identification is the customer's; third-party usage is recorded under MANAGE 3.1. |
| MAP 5.1 | Likelihood and magnitude of impacts | Impact assessments are recorded per asset in the Registry; this pack does not attach them as evidence. |
| MAP 5.2 | Engagement with AI actors on impacts | Same as MAP 5.1. |
| MEASURE 1.1 | Metrics for enumerated risks selected | Metric selection is the customer's. |
| MEASURE 1.2 | Appropriateness of metrics assessed | Same as MEASURE 1.1. |
| MEASURE 1.3 | Independent assessors involved | Assessor involvement is outside the product. |
| MEASURE 2.1 | Test sets, metrics and TEVV tools documented | Wytness does not test or evaluate the model. |
| MEASURE 2.2 | Human-subject evaluations | Same as MEASURE 2.1. |
| MEASURE 2.3 | Performance measured for deployment conditions | Same as MEASURE 2.1. |
| MEASURE 2.5 | Validity and reliability demonstrated | Model validation is the customer's. |
| MEASURE 2.6 | Safety risks evaluated | Safety evaluation is the customer's. |
| MEASURE 2.9 | Model explained, validated and documented | Explainability is outside Wytness's view. |
| MEASURE 2.11 | Fairness and bias evaluated | Wytness does not measure bias. |
| MEASURE 2.12 | Environmental impact assessed | Outside the product. |
| MEASURE 2.13 | Effectiveness of TEVV metrics evaluated | Wytness does not test or evaluate the model. |
| MEASURE 3.1 | Risks tracked over time | A risk register is the customer's; the anomaly history that feeds it is evidenced under MEASURE 2.4. |
| MEASURE 3.2 | Risk tracking where metrics are unavailable | Same as MEASURE 3.1. |
| MEASURE 3.3 | End-user feedback and appeal processes | A feedback and appeal channel; Wytness has none. |
| MEASURE 4.1 | Measurement approaches informed by domain experts | Consultation is the customer's. |
| MEASURE 4.2 | Trustworthiness results validated with AI actors | Same as MEASURE 4.1. |
| MEASURE 4.3 | Performance improvements or declines identified | Same as MEASURE 4.1. |
| MANAGE 1.2 | Risk treatment prioritised | Risk treatment is the customer's. |
| MANAGE 1.3 | Responses to high-priority risks planned | Same as MANAGE 1.2. |
| MANAGE 1.4 | Negative residual risks documented | Same as MANAGE 1.2. |
| MANAGE 2.1 | Resources and non-AI alternatives considered | A planning judgement; no Wytness mechanism. |
| MANAGE 2.2 | Mechanisms to sustain the value of deployed systems | Maintenance is the customer's. |
| MANAGE 2.3 | Procedures to respond to and recover from unknown risks | Incident response and recovery procedures are the customer's. |
| MANAGE 2.4 | Mechanisms to supersede, disengage or deactivate | `POST /agents/{agent_id}/kill` records the intent and attempt history only; the deactivation needs a handler in the customer's runtime (`api/app/models_pg.py`, `kill_events`). |
| MANAGE 3.2 | Pre-trained models monitored | The per-model call record is the MANAGE 3.1 row; model monitoring beyond usage is the customer's. |
| MANAGE 4.2 | Continual improvement activities | An improvement programme; no Wytness mechanism. |
| MANAGE 4.3 | Incidents and errors communicated | Communication to affected parties is the customer's process. |
- —The RMF is voluntary and has no certification. Wytness is not an assessor, and an Evidence Pack is a subcategory mapping with evidence, not a conformity statement.
- —Risk tolerance, impact assessment (MAP 5), risk treatment and residual risk (MANAGE 1.2 to 1.4) are judgements your organisation makes. We record what agents did; what that means for your risk posture is your assessment.
- —Testing, evaluation, validity, safety, explainability and bias measurement (MEASURE 2.1 to 2.3, 2.5, 2.6, 2.9, 2.11) are outside Wytness's view. We log operational inputs and outputs; we do not evaluate the model.
- —Incident response, recovery and the deactivation of an agent (MANAGE 2.3, 2.4, 4.3) belong to your runtime and your team. The kill endpoint records the attempt; it does not stop the agent.