The NIST AI Risk Management Framework (AI RMF 1.0, published as NIST AI 100-1) is a voluntary guidance document NIST released on 26 January 2023, after a direction from Congress, to "help manage the many risks of AI technologies." NIST Director Laurie E. Locascio described its goal as helping "companies and other organizations in any sector and any size to jump-start or enhance their AI risk management approaches." Unlike a regulation, the AI RMF sets out no mandatory controls or audits — it is a structure organizations can adopt, adapt, or borrow from selectively, built around four functions and seven characteristics of what NIST calls "trustworthy" AI.
Key takeaways
- NIST's AI Resource Center describes the Core as the framework's operational component: outcomes and actions grouped into four functions, each broken into categories and further into subcategories.
- Valid and reliable — performs as intended, consistently, under expected conditions.
- A Profile is how an organization applies the Core's functions, categories and subcategories to an actual setting.
- The Core has 72 subcategories in total.
What are the four functions of the AI RMF Core?
NIST's AI Resource Center describes the Core as the framework's operational component: outcomes and actions grouped into four functions, each broken into categories and further into subcategories.
| Function | Categories / subcategories | What it does |
|---|---|---|
| Govern | 6 / 19 | Cross-cutting function "infused throughout AI risk management" — policies, accountability structures, and culture that the other three functions rely on |
| Map | 5 / 18 | "Establishes the context to frame risks related to an AI system" — what the system is, who it affects, what it's for |
| Measure | 4 / 22 | "Employs quantitative, qualitative, or mixed-method tools, techniques, and methodologies to analyze, assess, benchmark, and monitor AI risk and related impacts" |
| Manage | 4 / 13 | "Entails allocating risk resources to mapped and measured risks on a regular basis and as defined by the govern function" |
NIST is explicit that the actions inside each function are "neither a checklist nor set of steps to be followed in its entirety" (the same language NIST uses to describe the companion Playbook, below). Governance is meant to run underneath the whole process; once it's in place, most organizations start with Map and move into Measure or Manage as the AI system's lifecycle warrants — not necessarily in a fixed order.
Seven characteristics of trustworthy AI
The AI RMF ties all four functions back to seven characteristics it defines as what "trustworthy" means for an AI system in practice:
- Valid and reliable — performs as intended, consistently, under expected conditions.
- Safe — should "not, under defined conditions, lead to a state in which human life, health, property, or the environment is endangered."
- Secure and resilient — maintains confidentiality, integrity and availability, and degrades safely under adverse events rather than failing catastrophically.
- Accountable and transparent — information about the system and its outputs is available to the people interacting with it.
- Explainable and interpretable — operators and affected people can understand how the system works and why it produced a given output.
- Privacy-enhanced — limits intrusion into individuals' autonomy, identity and data through norms and technical practice.
- Fair, with harmful bias managed — addresses harmful bias and discrimination, while recognizing fairness standards "differ across cultures and contexts."
NIST frames these as interrelated, not independent checkboxes: a decision that improves one (say, raw performance) can trade off against another (say, explainability), and weighing that trade-off throughout the AI lifecycle is explicitly part of what the Manage function is for.
Profiles: turning the Core into something concrete
A Profile is how an organization applies the Core's functions, categories and subcategories to an actual setting. NIST's AI Resource Center names three kinds: use-case Profiles for a specific sector or application, temporal Profiles contrasting a Current Profile (how AI risk is managed today) against a Target Profile (the desired state), and cross-sectoral Profiles for models or applications used across many sectors at once. The best-known published example is the Generative AI Profile (NIST AI 600-1, released 26 July 2024), built to "identify unique risks posed by generative AI" inside the same Govern/Map/Measure/Manage structure rather than inventing a parallel framework.
A companion Playbook translates the Core's subcategories into suggested, non-mandatory actions — NIST describes it as "neither a checklist nor set of steps to be followed in its entirety," meant to be borrowed from selectively rather than implemented wholesale.
Which AI RMF outcomes matter most for a model inside a device?
The Core has 72 subcategories in total. For a model that ships in firmware, a small set does most of the work, because the model is frozen into an image, runs on hardware you no longer control, and often comes from a supplier. The wording below is quoted from the subcategory text in NIST AI 100-1.
| Subcategory | NIST text (abridged) | Device question | Evidence to keep |
|---|---|---|---|
| GOVERN 6.1 | "Policies and procedures are in place that address AI risks associated with third-party entities" | Who supplied the model, the runtime and the training data, and under what terms? | Supplier list, licences, contract clauses on updates and disclosure |
| MAP 4.1 | Approaches for mapping risks of AI components, "including the use of third-party data or software", are "in place, followed, and documented" | Which components make up the AI feature (weights, runtime library, pre- and post-processing code)? | A component inventory, for example a CycloneDX ML-BOM alongside the firmware SBOM |
| MEASURE 2.4 | "The functionality and behavior of the AI system and its components … are monitored when in production" | How do you see accuracy drift on devices in the field? | Telemetry design, sampled field-error reports |
| MEASURE 2.7 | "AI system security and resilience … are evaluated and documented" | Can the model file be swapped, extracted or fed adversarial input? | Test reports, signing and secure-boot design |
| MANAGE 2.4 | Mechanisms to "supersede, disengage, or deactivate AI systems" that perform outside intended use | Can you switch the AI feature off or roll the model back remotely? | Feature-flag and rollback design, tested |
| MANAGE 3.2 | "Pre-trained models which are used for development are monitored as part of AI system regular monitoring and maintenance" | Do you track new issues in the base model and its runtime? | Watch process for the base model and runtime library |
| MANAGE 4.1 | "Post-deployment AI system monitoring plans are implemented", including user input, appeal and override, and decommissioning | What happens when a user reports a wrong decision, and at end of support? | Support process, decommissioning plan |
Most of these overlap with work a device maker already does for conventional firmware: a supplier register, an SBOM, a vulnerability-handling process and a secure update path. The AI RMF adds the model-specific questions on top: behaviour in the field, the ability to disengage, and the base model as a monitored dependency.
How do you apply the AI RMF to an edge-AI product, step by step?
NIST's Playbook stays non-prescriptive, so the order below is one workable sequence for a product team, not a NIST procedure.
Set the governance owner (Govern)Name who decides whether an AI feature ships, who can disable it, and who owns supplier relationships for the model and runtime. Describe the system and its context (Map)Write down intended use, users, deployment environment (offline, intermittent or always connected), and what a wrong output can cause: a missed alert, a false alarm, a safety function triggered or withheld. Inventory the AI components (Map)List the model file and its version and hash, the inference runtime, its own dependencies, and the origin of the base model and training data. Put this next to the firmware component list so both are versioned per release. Choose what to measure (Measure)Pick tests that match the context: accuracy on data from the real deployment environment, robustness to the input manipulations in the NIST AI 100-2 E2025 taxonomy that apply to your sensor, and integrity checks on the model file. Decide the controls (Manage)Typical choices: sign the model and verify it at build and boot, keep a remote disable switch, set a support period for the model as well as the firmware. Write the Current and Target ProfileRecord which subcategories are met today, which are targeted for the next release, and who owns each gap. Re-run Map and Measure on every model updateA retrained model is a new component; treat it like a firmware release.
Worked example: a Current and Target Profile for a smart camera
Synthetic example. The company, product and figures are invented for illustration.
A hypothetical maker ships an outdoor camera with an on-device person-detection model licensed from a supplier. The firmware is updated over the air about four times a year. The team writes a temporal Profile covering five subcategories:
| Subcategory | Current | Target (next release) | Owner |
|---|---|---|---|
| GOVERN 6.1 | Licence signed; no clause on model updates | Contract clause: supplier notifies model issues within 30 days | Purchasing |
| MAP 4.1 | Firmware SBOM exists; model listed only by filename | Model, version, hash and runtime recorded per release | Firmware lead |
| MEASURE 2.7 | No integrity check on the model file | Model signature verified in CI; image covered by secure boot | Security lead |
| MANAGE 2.4 | Person detection cannot be disabled remotely | Cloud flag disables detection and falls back to motion alerts | Product owner |
| MANAGE 4.1 | Wrong detections go to general support | Support tags "AI decision" tickets and reviews them monthly | Support lead |
The value of the table is that each gap now has an owner and a release, and an auditor or customer can see the trajectory without the team claiming more than it does.
How does the AI RMF relate to other NIST AI documents?
- NIST AI 600-1, the Generative AI Profile (July 2024), is a cross-sectoral Profile of the AI RMF. It defines risks "unique to or exacerbated by" generative AI and maps suggested actions to Core subcategories. It applies if your device runs or calls a generative model. See our AI 600-1 guide.
- NIST AI 100-2 E2025 is a taxonomy of adversarial machine-learning attacks and mitigations. It feeds the Measure function for security and resilience. See our AI 100-2 guide.
- Model provenance sits under MEASURE 2.7 and MANAGE 3: signing the model file is one concrete control, covered in our guide to OpenSSF model signing.
- Inventory for MAP 4.1 is covered in our guide to the AI-BOM for embedded AI devices.
For products sold into the EU, the AI RMF is voluntary and does not replace legal duties; see our guide to EU AI Act cybersecurity requirements.
What are the common pitfalls?
- Treating the model as data, not a component. If the model file has no version, hash and owner, MAP 4.1 and MANAGE 3.2 cannot be met, and a swapped file goes unnoticed.
- Measuring only lab accuracy. Edge models meet lighting, noise, temperature and sensor ageing that lab test sets miss; MEASURE 2.4 asks for monitoring in production.
- No off switch. A model that misbehaves in the field and cannot be disabled without a full firmware release fails MANAGE 2.4.
- Forgetting the runtime. The inference library is ordinary software with its own CVEs and belongs in the firmware SBOM.
- Writing a Profile once. A Profile that is not updated when the model is retrained describes a product that no longer exists.
- Claiming "AI RMF compliant". The framework is voluntary and has no certification; describe which subcategories you address instead.
Where this fits for device and firmware fleets
Products that embed an on-device model, call out to a cloud model, or expose an AI feature through firmware carry AI-specific risk on top of conventional software risk — model behavior drift, data used for inference, and explainability of an automated decision are not questions a traditional vulnerability scan answers. Teams already mapping conventional application risk against OWASP's LLM-specific list (see our OWASP Top 10 for LLM Applications guide) or securing an MCP integration (see MCP server security) can use the AI RMF's four functions as the governance layer that sits above both: Govern sets the policy, Map identifies which device features actually carry AI risk, Measure benchmarks that risk concretely, and Manage decides what to do about it.
For the conventional part of that surface, SoftWSP extracts a CycloneDX 1.6 SBOM from a firmware image, matches its components (including the inference runtime and its libraries) against NVD, CISA KEV, FIRST EPSS and CISA ICS advisories, and drafts a cited report for engineer review. Model behaviour testing stays with your AI team. See device & firmware security services.
This guide summarises a public NIST publication; it does not constitute legal or compliance advice.
Put this guide to work
Frequently asked questions
What is the NIST AI Risk Management Framework?
A voluntary guidance document published by NIST on 26 January 2023 as NIST AI 100-1, developed following a direction from Congress. NIST describes it as intended to "help manage the many risks of AI technologies" and to "help companies and other organizations in any sector and any size to jump-start or enhance their AI risk management approaches" (NIST Director Laurie E. Locascio, quoted in NIST's own announcement). It is explicitly non-mandatory: NIST calls it a "guidance document for voluntary use by organizations."
What are the four core functions — Govern, Map, Measure, Manage?
NIST's AI Resource Center describes the AI RMF Core as organized into four functions, each broken into categories and subcategories: GOVERN (6 categories, 19 subcategories) is described as a "cross-cutting function that is infused throughout AI risk management and enables the other functions"; MAP (5 categories, 18 subcategories) "establishes the context to frame risks related to an AI system"; MEASURE (4 categories, 22 subcategories) "employs quantitative, qualitative, or mixed-method tools, techniques, and methodologies to analyze, assess, benchmark, and monitor AI risk and related impacts"; and MANAGE (4 categories, 13 subcategories) "entails allocating risk resources to mapped and measured risks on a regular basis and as defined by the govern function." NIST states the actions are not a checklist or a mandatory sequence — once governance is in place, most users start with Map and continue into Measure or Manage as needed.
What are the seven characteristics of trustworthy AI?
Per the AI RMF's own characteristics section: valid and reliable, safe (an AI system should "not, under defined conditions, lead to a state in which human life, health, property, or the environment is endangered"), secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair — with harmful bias managed. NIST notes fairness standards "differ across cultures and contexts," and frames these seven as interrelated rather than independent checkboxes — optimizing one (for example performance) can trade off against another (for example transparency), which is itself something the Manage function is meant to surface and weigh.
What is an AI RMF Profile, and how does the Generative AI Profile fit in?
A Profile is NIST's term for "implementations of the AI RMF functions, categories, and subcategories for a specific setting or application based on the requirements, risk tolerance, and resources of the Framework user." NIST's AI Resource Center describes three kinds: use-case Profiles (for a specific sector, technology or application, such as hiring), temporal Profiles (a Current Profile describing how AI is managed today versus a Target Profile describing the desired state), and cross-sectoral Profiles (covering models or applications used across many sectors). The best-known published example is the Generative AI Profile, NIST AI 600-1, released 26 July 2024 to help organizations "identify unique risks posed by generative AI" within the same four-function structure.
Sources
- NIST: AI Risk Management Framework (overview, four functions, related resources)
- NIST AI Resource Center: AI RMF Core — Govern/Map/Measure/Manage definitions and category/subcategory counts
- NIST AI Resource Center: Characteristics of Trustworthy AI Systems
- NIST AI Resource Center: AI RMF Profiles
- NIST AI Resource Center: AI RMF Playbook
- NIST News: NIST Risk Management Framework Aims to Improve Trustworthiness of Artificial Intelligence (26 January 2023 — publication date, voluntary status, Congressional direction)
- NIST AI 100-1: Artificial Intelligence Risk Management Framework (AI RMF 1.0), PDF (Core subcategory text)
- NIST AI 600-1: Generative Artificial Intelligence Profile (July 2024), PDF
- NIST CSRC: AI 100-2 E2025, Adversarial Machine Learning taxonomy
- CycloneDX: Machine Learning Bill of Materials (ML-BOM)
About this guide
SoftWSP Regulation Desk
Covers the Cyber Resilience Act cluster: 34 guides.

In this cluster
Questions about this page?
