NIST AI Risk Management Framework
AI RMF 1.0 (NIST AI 100-1, January 2023)
This is not an assessment. These mappings indicate which ARC Control evidence is relevant to a framework reference. They are not an assessment, not a certification, and not a statement that the reference is satisfied. The AI RMF is voluntary guidance, not a control standard with pass criteria; any conclusion about conformity is the organization’s own to reach. References are abbreviated for display and should be checked against the published framework text.
17 references mapped0 reviewed0 evidenced
| Reference | Review | Controls | Evidenced | What ARC does and does not evidence |
|---|---|---|---|---|
| GOVERN 1.3 Risk management activities are scaled to the organization’s risk tolerance | proposed | ARC-AI-002 <span class="badge allow">Pass</span> ARC-AI-008 <span class="badge deny">Fail</span> | no | ARC evidences that high-impact tools are deliberately governed and that agent permissions are periodically reviewed. It does not evidence that the organization has set a risk tolerance, nor that the scaling matches it. Gap: the mapping has not been reviewed |
| GOVERN 1.5 Ongoing monitoring and periodic review, with defined roles and review frequency | proposed | ARC-AI-008 <span class="badge deny">Fail</span> | no | ARC records review cadence per agent and reports overdue reviews. The adequacy of the cadence, and whether reviewers are competent, are outside what ARC can evidence. Gap: the mapping has not been reviewed |
| GOVERN 2.1 Roles, responsibilities and accountability are documented | proposed | ARC-AI-001 <span class="badge deny">Fail</span> | no | ARC evidences a named accountable owner per agent. It cannot evidence that the named person accepts the accountability or is still in post. Gap: the mapping has not been reviewed |
| GOVERN 3.2 Human oversight roles for AI systems are defined and differentiated | proposed | ARC-AI-004 <span class="badge allow">Pass</span> | no | ARC evidences that policy-required human approvals occurred before execution, with separation of duties where demanded. In this release approver identity is a configured header rather than an authenticated user, which materially limits what the evidence establishes about *who* exercised oversight. Gap: the mapping has not been reviewed |
| GOVERN 4.3 Practices for reporting and documenting AI risks are in place | proposed | ARC-AI-005 <span class="badge allow">Pass</span> ARC-AI-006 <span class="badge info">Not assessable</span> | no | ARC evidences that attempts and refusals are recorded and that the record is tamper-evident. ARC-AI-006 does not pass without an externally anchored audit head. Gap: the mapping has not been reviewed |
| GOVERN 6.1 Policies address risks from third-party software and data | proposed | ARC-AI-007 <span class="badge muted">Not applicable</span> | no | ARC evidences that MCP tool manifests are pinned so upstream change is detectable. It does not assess the third party itself, and vendor assessment is not implemented in this release. Gap: the mapping has not been reviewed |
| MAP 1.1 Intended purpose and context of use are understood and documented | proposed | ARC-AI-001 <span class="badge deny">Fail</span> | no | ARC stores a stated business purpose and criticality per agent. The text is free-form and ARC cannot judge its accuracy or currency. Gap: the mapping has not been reviewed |
| MAP 2.1 The tasks the system supports are defined | proposed | ARC-AI-001 <span class="badge deny">Fail</span> ARC-AI-002 <span class="badge allow">Pass</span> | no | The tool registry and the policies selecting it together document what an agent is permitted to attempt, which is a narrower and more reliable statement than what it is intended to do. Gap: the mapping has not been reviewed |
| MAP 4.1 Risks of third-party components are mapped | proposed | ARC-AI-007 <span class="badge muted">Not applicable</span> ARC-AI-009 <span class="badge approval">Attested</span> | no | Manifest pinning covers MCP components. ARC-AI-009 covers whether a governed system is reachable outside ARC, and it is attestation-based: ARC cannot test it. Gap: the mapping has not been reviewed |
| MEASURE 1.1 Approaches and metrics for measuring AI risk are selected | proposed | ARC-AI-002 <span class="badge allow">Pass</span> ARC-AI-003 <span class="badge allow">Pass</span> ARC-AI-004 <span class="badge allow">Pass</span> ARC-AI-005 <span class="badge allow">Pass</span> | no | The control library is itself the selected approach for the authorization and evidence risks. It deliberately does not address model-level risks such as accuracy, bias or drift, which ARC does not observe. Gap: the mapping has not been reviewed |
| MEASURE 2.7 Security and resilience are evaluated and documented | proposed | ARC-AI-003 <span class="badge allow">Pass</span> ARC-AI-006 <span class="badge info">Not assessable</span> ARC-AI-007 <span class="badge muted">Not applicable</span> ARC-AI-009 <span class="badge approval">Attested</span> | no | Covers enforcement, log integrity, supply-chain drift detection and bypass prevention. It is not a penetration test and makes no statement about resilience under load or attack. Gap: the mapping has not been reviewed |
| MEASURE 2.8 Transparency and accountability risks are examined | proposed | ARC-AI-005 <span class="badge allow">Pass</span> ARC-AI-006 <span class="badge info">Not assessable</span> | no | Every decision carries a rule-by-rule trace and an evidence record, and the record is exportable and independently checkable. The limits of tamper-evidence are stated on the control rather than omitted. Gap: the mapping has not been reviewed |
| MEASURE 3.1 Mechanisms exist to determine whether assurance criteria are met | proposed | ARC-AI-001 <span class="badge deny">Fail</span> ARC-AI-002 <span class="badge allow">Pass</span> ARC-AI-003 <span class="badge allow">Pass</span> ARC-AI-004 <span class="badge allow">Pass</span> ARC-AI-005 <span class="badge allow">Pass</span> ARC-AI-006 <span class="badge info">Not assessable</span> ARC-AI-007 <span class="badge muted">Not applicable</span> ARC-AI-008 <span class="badge deny">Fail</span> | no | The assessment runner is this mechanism. Its honest limit is recorded in the results themselves: a control ARC cannot decide reports not assessable rather than passing. Gap: the mapping has not been reviewed |
| MANAGE 1.2 Treatment of documented risks is prioritized by impact and likelihood | proposed | none mapped | no | Addressed by the risk register rather than by a control: risks carry inherent and reviewed-residual ratings, a treatment decision and an owner. No ARC control evidences this, and the mapping says so rather than borrowing a control that does not fit. Gap: the mapping has not been reviewed |
| MANAGE 2.2 Mechanisms sustain the value of deployed systems | proposed | ARC-AI-008 <span class="badge deny">Fail</span> | no | Periodic permission review is the only part of this ARC speaks to. Most of the reference is about model value and performance, which ARC does not observe. Gap: the mapping has not been reviewed |
| MANAGE 2.4 Mechanisms exist to deactivate systems behaving inconsistently with intended use | proposed | ARC-AI-003 <span class="badge allow">Pass</span> | no | ARC provides agent suspension, tool and connector disablement and retirement, each audited. The in-flight limitation is documented: disabling stops new actions and does not recall one already executing. Gap: the mapping has not been reviewed |
| MANAGE 4.1 Post-deployment monitoring plans are implemented | proposed | ARC-AI-003 <span class="badge allow">Pass</span> ARC-AI-005 <span class="badge allow">Pass</span> | no | Continuous evaluation over a stated window, with the window on every result. ARC monitors authorization and evidence, not model behaviour. Gap: the mapping has not been reviewed |