Zulia Shavaeva← Back to Zulia
Free AI and security guide/Learn

Learn / AI governance, risk and compliance: NIST AI RMF, ISO/IEC 42001, EU AI Act, CSA AICM

AI governance, risk and compliance: NIST AI RMF, ISO/IEC 42001, EU AI Act, CSA AICM

Applying formal risk management, management-system, and regulatory frameworks to how an organisation builds, buys, and operates AI.

Researched on 2026-09-26 with AI assistance. Links and summaries can change; verify details with the original source. Not yet reviewed by a person.

What it is

This skill is about running AI risk through formal structures: a voluntary risk framework (NIST AI RMF 1.0), a certifiable management system (ISO/IEC 42001), a binding regulation (EU AI Act), and a control set you can audit against (CSA AI Controls Matrix). NIST frames AI risk as socio-technical, arguing that AI risks differ from traditional software risks because training data shifts over time and failures are hard to detect in deployment.1

The regulatory anchor in Europe is Regulation (EU) 2024/1689, published in the Official Journal on 12 July 2024, which sets risk-based rules for AI providers and deployers rather than rules about AI techniques.106

Frameworks are complementary, not competing, and the practical work is crosswalking them: CSA's AI Controls Matrix v1.1 ships mappings to ISO 42001:2023, ISO 27001, BSI AIC4, the EU AI Act, and NIST AI RMF plus AI 600-1, so one control can satisfy several obligations.8

Day to day it looks like inventories, risk registers, classification decisions, control design, evidence collection and third-party assessment questionnaires, applied to models and AI features instead of only to servers and applications.87

Why postings ask for it

14 of 48 postings ask for it (29%), and it is the defining skill of the governance cluster at 86% of 7 postings, so it is the entry requirement for those roles rather than a nice-to-have.P

Demand spreads into design and delivery work: 57% of 7 AI Security Architect postings and 40% of 5 consulting postings ask for it, which matches work like classifying a use case against the Act's risk tiers and choosing controls before build starts.P6

It is near-absent in red team (0% of 9) and research (0% of 4) and low in engineering (12% of 16), so treat it as the language you use to translate technical findings into obligations, evidence and owners.P

Concepts you should be able to explain

If you can say each of these out loud in two minutes, with an example, you are ready for the technical part of an interview on this skill.

NIST AI RMF 1.0 framing

The AI RMF is voluntary, rights-preserving, non-sector-specific and use-case agnostic, written for any organisation designing, developing, deploying or using AI. Part 1 frames AI risk and names the audience as AI actors, using the OECD definition of those who play an active role in the AI lifecycle. Part 1 also sets out characteristics of trustworthy AI systems, including valid and reliable, safe, and secure and resilient. Note that NIST says AI RMF 1.0 is being updated.1

Generative AI Profile (NIST AI 600-1)

The GenAI Profile is a cross-sectoral companion to the AI RMF for generative systems, published as NIST AI 600-1. Third-party documentation summarises it as 12 risk categories, including confabulation (false or misleading output) and CBRN information risk. Use it when your risk register needs GenAI-specific entries rather than generic model risk.23

ISO/IEC 42001 as a management system

ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system, is a certifiable AI management system standard published in 2023. Like other ISO management systems it separates mandatory clauses (context, leadership, planning, support, operation, performance evaluation, improvement) from a reference control set in Annex A; secondary guides summarise Annex A as 38 controls under nine objectives, and counts differ between commentators, so read the standard itself. The point for an interview is that 42001 governs how the organisation manages AI, not whether one model is safe.45

EU AI Act risk tiers and prohibitions

The Act sets four levels of risk and bans nine practices outright, including social scoring, untargeted scraping to build facial recognition databases, emotion recognition in workplaces and schools, and real-time remote biometric identification for law enforcement in public spaces. High-risk use cases include credit scoring, CV sorting and worker management, education scoring, biometric identification, law enforcement, migration, and safety components of critical infrastructure and products. Classification is driven by use case and context, so the same model can be high-risk in one product and not in another.6

Provider and deployer roles

Obligations attach to roles. A provider develops an AI system or general-purpose AI model, or has one developed, and places it on the market or puts it into service under its own name, whether paid or free; deployers use systems supplied by others. The Act's definition of an AI system covers machine-based systems that operate with varying autonomy, may adapt after deployment, and infer outputs such as predictions, content, recommendations or decisions. Fine-tuning and rebranding a model can move you from deployer to provider, which changes the paperwork you owe.76

Article 9 risk management system

For high-risk systems, Article 9 requires a risk management system that is established, implemented, documented and maintained, and treated as a continuous iterative process across the whole lifecycle with regular systematic review and updating. It requires identification and analysis of known and reasonably foreseeable risks to health, safety and fundamental rights when the system is used for its intended purpose, then estimation and evaluation of those risks. This is where a familiar ISO 27005 style risk process transfers, with fundamental rights added to the impact axis.7

Phased application and amendment

Obligations arrive in stages: prohibitions 1 to 8 became effective in February 2025, the ninth prohibition (AI-generated non-consensual intimate imagery and CSAM) applies from December 2026 and was introduced through the AI Omnibus, and high-risk obligations start from 2 December 2027. Article 9 itself is listed as coming into force on 2 December 2027 for Annex III high-risk systems and 2 August 2028 for Annex I systems under Article 113(c). Because the text is being amended, always check the date and version you are advising on.67

CSA AICM and the AI-CAIQ

The AI Controls Matrix v1.1 is a free vendor-agnostic control set for cloud-based AI: 247 control objectives across 18 security domains, analysed by control type, applicability and ownership, architectural relevance, LLM lifecycle relevance and threat category. It splits implementation and audit guidance by role (model provider, orchestrated service provider, application provider, AI customer, cloud service provider) and ships the AI-CAIQ questionnaire for self-assessment or vendor review, plus a STAR for AI Level 1 submission guide and machine-readable JSON/YAML/OSCAL bundles. It is the most direct bridge from framework text to a control list you can test.8

Use fictional data and authorised sandboxes. Remove employer details and secrets from any portfolio write-up. Time estimates exclude setup. Check model and cloud costs before running tests, set spending limits, and delete lab resources afterwards.

Three exercises

In order of difficulty. Free tools. Keep what you build; it is evidence.

1Classify three AI use cases against the EU AI Actabout 3 h

You can state, with article references, whether a use case is prohibited, high-risk or neither, and whether your organisation is provider or deployer.67

  1. Pick three real use cases, for example a CV-screening assistant, an internal RAG helpdesk bot, and an emotion analytics pilot for staff calls.
  2. Check each against the nine prohibited practices and the high-risk use case list on the Commission's AI Act page.
  3. Decide provider or deployer for each using the Article 3 provider definition, and note what changes if you fine-tune a vendor model.
  4. Write one page per use case: tier, reasoning, role, the date obligations begin, and what you would need to confirm with legal.

Tools: web browser, text editor or spreadsheet

2Build an Article 9 style risk register for a GenAI featureabout 6 h

A risk register a reviewer accepts: named risks traced to a published taxonomy, impact on health, safety and fundamental rights, and review triggers.7921

  1. Choose one GenAI feature and write its intended purpose and conditions of use in plain language.
  2. Pull candidate risks from the MIT AI Risk Repository domain taxonomy (7 domains, 24 subdomains) and from the GenAI Profile risk categories such as confabulation.
  3. For each risk record the Article 9 steps: identification and analysis, then estimation and evaluation under intended use.
  4. Add the socio-technical angle NIST stresses: who operates it, who is affected, and how a failure would be detected.
  5. Define the review trigger and cadence that makes the register a continuous iterative process rather than a one-off document.

Tools: MIT AI Risk Repository database copy, spreadsheet

3Run an AICM gap assessment with crosswalksabout 10 h

A gap report that shows control coverage for one AI service, evidence per control, and which ISO 42001 and EU AI Act obligations each gap touches.8

  1. Download AICM v1.1 and pick the role view that matches your position (for example AI Customer or Application Provider).
  2. Scope to one AI service and select the relevant domains from the 18, filtering by LLM lifecycle relevance and threat category.
  3. Complete the AI-CAIQ for that scope, recording owner and evidence for each answer rather than yes or no.
  4. Use the bundled mappings to ISO 42001:2023, the EU AI Act and NIST AI RMF and 600-1 to show which obligations each gap affects.
  5. Write a two-page report: top five gaps, owners, proposed control text, and what evidence would close each gap.

Tools: CSA AICM v1.1 spreadsheet and AI-CAIQ, spreadsheet

Practice questions

Written from the concepts above, not collected from a named employer. Open one, answer it out loud, then tick the points you covered; the score stays in this browser.

Why does NIST argue AI risk needs its own framework rather than reuse of existing software risk practice?NIST AI RMF 1.0 framing

Say your answer out loud or write it down, then tick what you covered:

0 of 4 covered

A product team wants to fine-tune a vendor model and ship it under our brand. What changes from a compliance point of view?Provider and deployer roles

Say your answer out loud or write it down, then tick what you covered:

0 of 4 covered

Walk me through what Article 9 actually requires for a high-risk system.Article 9 risk management system

Say your answer out loud or write it down, then tick what you covered:

0 of 4 covered

What is the difference between certifying against ISO/IEC 42001 and conforming to the EU AI Act?ISO/IEC 42001 as a management system

Say your answer out loud or write it down, then tick what you covered:

0 of 4 covered

How would you choose controls for a customer-facing GenAI assistant without inventing a framework?CSA AICM and the AI-CAIQ

Say your answer out loud or write it down, then tick what you covered:

0 of 4 covered

Which AI use cases are simply off the table in the EU, and which are heavily regulated?EU AI Act risk tiers and prohibitions

Say your answer out loud or write it down, then tick what you covered:

0 of 4 covered

Where do you get your risk list from, and how do you avoid a register that is just a list of buzzwords?MIT AI Risk Repository and Generative AI Profile

Say your answer out loud or write it down, then tick what you covered:

0 of 4 covered

How do you keep advice current when the frameworks themselves are moving?Phased application and amendment

Say your answer out loud or write it down, then tick what you covered:

0 of 4 covered

Sources

Every numbered claim above links here. P = the platform's own coding of 48 job postings.

  1. AI Risk Management Framework 1.0, Executive Summary (AIRC) NIST
  2. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1) NIST
  3. NIST AI RMF Generative AI Profile (NIST AI 600-1), 12 risk categories and operationalization (secondary summary, not a NIST page) Modulos documentation
  4. ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system (catalogue entry; page returned HTTP 403 on fetch, details from the search listing; price and page count not stated) ISO/IEC
  5. ISO 42001 Requirements: Clauses and Controls (secondary guide; Annex A control counts differ between commentators) Compyl
  6. AI Act: regulatory framework for AI European Commission
  7. Article 9: Risk Management System, EU AI Act Explorer artificialintelligenceact.eu
  8. AI Controls Matrix (AICM) v1.1 Cloud Security Alliance
  9. AI Risk Repository, risk database and taxonomies MIT AI Risk Initiative
  10. The Act Texts (Official Journal publication date and text versions) artificialintelligenceact.eu

Resources

Free first. Levels: intro means no prior knowledge of this skill; working means you can apply it on a project; advanced means research depth or specialist tooling.

Level Format

Gaps the research could not fill with a good free source: ISO/IEC 42001 itself: the ISO catalogue page did not appear in any search result today, so the existing entry was dropped rather than reproduced from memory. Also found no free full text (ISO standards are paid) and no free implementation course.; ISO/IEC 23894 (AI risk management guidance) and the 42001 to NIST AI RMF crosswalk: no confirmed source page.; An article-by-article EU AI Act explorer: the artificialintelligenceact.eu URL in the existing list did not reappear in a search result, so official Commission pages are used instead.; NIST AI 100-2 (adversarial machine learning taxonomy): no confirmed URL in today's results, although GRC readers need it for control mapping.; A free hands-on conformity assessment or high-risk classification lab for the EU AI Act (fill-in exercise with model answers).; A free ISO 42001 or AICM internal audit workbook with worked evidence examples; CSA and IAPP audit training in this space is paid.; Confirmed course durations and prices for CSA STAR AI Controls Auditor training: page not fetched, price not stated in search results.