NIST AI RMF with ARRTECH
Records for the AI systems in a risk register.
Records for AI in security operations
AI risk management needs a clear account of the system’s role, evidence and authority. Evaluation results and recorded review decisions make those responsibilities practical to examine.
Requirements
| Clause | What it asks | Security support |
|---|---|---|
| GOVERN 1.6 | An AI system inventory. | Each detector’s state. → |
| MAP 3.5 | Human oversight defined. | Who accepted each case. → |
| MEASURE 2.5 | Validity and reliability shown. | Measured in shadow before promotion. → |
| MANAGE 2.4 | Systems superseded or disengaged. | Every rollback on record. → |
| MANAGE 4.1 | Post-deployment monitoring. | Data health on each finding. → |
Inventory
An inventory identifies each detector, its purpose and whether it is disabled, being tested in shadow or operating in production. Keep the record aligned with the deployed configuration.
Review
A case should present its assessment, supporting evidence and unresolved questions for review. Record who accepted, challenged or returned the work and the reason for that decision.
Measurement
In shadow, replay measures a detector’s precision, recall, stability and volume before it goes live.
Rollback
Any AI Detection detector can go back to shadow or be disabled, and the change is recorded.
Residency
Local inference can keep analytical processing within the organization’s environment. Any external model route is an explicit policy choice, with permitted data and access defined.
Limits
The following responsibilities belong to the organization’s wider security and governance program. Their owners maintain the evidence needed for review.
MAP 1.1
Intended context of use.
MAP 5.1
Likelihood and size of impacts.
GOVERN 6.1
Third-party AI review needs the documentation relevant to the models and services used. Record the sources, responsibilities and information still needed for the proposed use.
Data provenance
Document the data sources and permitted uses relevant to the model and deployment. Distinguish local behavioral learning from language-model inference and any external data processing.
By rule
Use these references to connect the requirement with relevant operating evidence. The organization defines the scope and implementation being assessed.
GOVERN 1.6
Maintain an inventory of AI systems. An inventory identifies the detector, its role and its current operating state.
MAP 3.5
Define and document processes for human oversight. A SOAR incident workflow retains ownership, evidence, approvals and actions so reviewers can follow the response.
MEASURE 2.5
Demonstrate that the system is valid and reliable. Shadow evaluation and recorded release decisions show how the detector was assessed before production use.
MANAGE 2.4
Mechanisms to supersede, disengage or deactivate AI systems. Recorded rollback or deactivation lets the operating team withdraw a capability and retain the reason.
MANAGE 4.1
Monitor AI systems after deployment. Source-health information helps reviewers judge whether a result rests on usable and sufficiently complete data.
Questions
Is the AI RMF mandatory?
No. NIST intends it for voluntary use.
Does inference leave our environment?
Not by default.
Is AI Investigation available?
In early access. It investigates an AI Detection finding or a SIEM alert.
Who accepts an Investigation case?
An analyst. The case records who accepted it.
Guide
What is the NIST AI RMF?
A voluntary NIST framework for managing AI risk, organized as Govern, Map, Measure and Manage.
What is AI 600-1?
The NIST profile that applies the AI RMF to generative AI, published in July 2024.
Is there an AI RMF certification?
NIST does not offer an AI RMF certification. The framework is intended to help organizations manage and communicate AI risks.
How ARRTECH takes care of NIST AI RMF records
Connects security activity into a signed record, giving incident decisions and board reporting evidence that can be independently verified.
Protects intellectual property and customer trust, with content-aware controls that let organizations adopt AI on their own terms.
Expands response capacity through automation, with defined human authority and a record of who approved what.
Finds changes in behavior that predefined rules miss, connecting the signs of an attack so teams can understand what warrants action.
Investigates alerts, tests explanations and prepares the evidence, giving analysts more capacity while people retain authority over the response.
