Catalog My learning Blog
AI Security Course: OWASP LLM Top 10 2026 and Agentic AI · Phase 6 · Lesson 28 · 4 min read

AI Governance: NIST AI RMF, MITRE ATLAS, ISO 42001

On this page
  1. Why engineers should care about AI governance frameworks
  2. NIST AI RMF, MITRE ATLAS and ISO/IEC 42001 compared
  3. Mapping your evidence to all three frameworks once
  4. What governance adds that engineering misses
  5. Check yourself

Why engineers should care about AI governance frameworks

Sooner or later someone asks: "How do we know our AI systems are secure?" That question comes from a CISO, an auditor, a big customer's security questionnaire, or a regulator. Answering "we added guardrails" doesn't work. Answering with a framework mapping does: here is each risk the framework names, here is our control, here is the test that proves it.

The good news is that you've already built most of the evidence. The risk register (Day 2), threat model (Day 4), CI gate (Day 24), report (Day 25) and monitoring (Day 27) map straight onto these frameworks. This lesson shows how.

NIST AI RMF, MITRE ATLAS and ISO/IEC 42001 compared

FrameworkWhat it isWhat it asks of you
NIST AI RMF 1.0 (2023)A voluntary risk-management frameworkGovern, Map, Measure and Manage AI risks
NIST AI 600-1 (2024)The Generative AI Profile of the RMFAddress 12 GenAI-specific risks with suggested actions
MITRE ATLASA knowledge base of real attacks on AI systemsName threats with shared technique ids
ISO/IEC 42001:2023A certifiable AI management system standardPolicies, roles, risk assessments, controls, audits

NIST AI RMF organizes work into four functions. Govern sets policies, roles and accountability. Map establishes the context, what the system does, who it affects and how. Measure assesses and tests the risks. Manage prioritizes and responds to them. It's voluntary in the US, but widely used as a common language.

NIST AI 600-1, the Generative AI Profile, lists 12 risks specific to generative AI, among them confabulation, data privacy, information security, information integrity, harmful bias, intellectual property, and value chain and component integration, plus suggested actions for each.

MITRE ATLAS is ATT&CK for AI systems. Version 5.1 (November 2025) lists 16 tactics and 84 techniques, including agent-specific ones such as tool credential harvesting. Use its ids to label threats and test cases, as you did in the threat model.

ISO/IEC 42001 is the one organizations get certified against. It's a management-system standard, like ISO 27001 for information security: it requires an AI policy, defined responsibilities, risk and impact assessments, controls, monitoring and continual improvement.

Mapping your evidence to all three frameworks once

Keep a single mapping in the repository that ties framework requirements to controls and evidence. Reviewers and auditors then read one file instead of interviewing ten engineers:

Code
# security/framework_map.yaml
- requirement: "NIST AI RMF MAP: context and risks are identified"
  evidence: [security/threat_model.yaml, security/risk_register.py]
- requirement: "NIST AI RMF MEASURE: risks are tested and tracked"
  evidence: [redteam/regressions.yaml, redteam/redteam.yaml, reports/2026-10-redteam.md]
- requirement: "NIST AI 600-1 Information Security: prompt injection and data exfiltration"
  controls: [reply_to_customer fixed recipient, quarantined reader, output renderer]
  atlas: [AML.T0051]
  tests: ["F-03 order-note exfiltration", "promptfoo owasp:llm"]
- requirement: "NIST AI 600-1 Confabulation"
  controls: [policy facts from tools, grounding check before high-stakes answers]
  tests: [evals/faithfulness.yaml]
- requirement: "NIST AI RMF MANAGE: incidents are handled"
  evidence: [runbooks/llm-incident.md, flags/kill-switches.md, drills/2026-09-drill.md]
- requirement: "ISO/IEC 42001: AI risk assessment is performed and reviewed"
  evidence: [security/risk_register.py, reviews/2026-q3-ai-risk-review.md]

A CI check can verify that every listed file exists and every named test is present in the suites, so the map can't quietly go stale.

What governance adds that engineering misses

Frameworks also ask questions engineers tend to skip:

  • Who owns each AI system, and who can approve a risky change such as a new tool or data source?
  • What's the inventory? Shadow AI, teams calling model APIs without review, is a real gap in most organizations.
  • Who is affected, beyond users? Employees whose work is scored, customers whose data trains a model.
  • When is a system retired, and what happens to its data and memories?

Answer those once, write them down, and the next questionnaire takes an hour instead of a month.

Check yourself

The risk register and threat model for identification, the CI red-team gate and report for testing, and the monitoring setup and incident playbook for response, organized by the NIST AI RMF functions.

The AI RMF is a voluntary framework and vocabulary for managing AI risk. ISO/IEC 42001 is a management-system standard an organization can be audited and certified against.

They give a shared, specific name for each technique, link to known real attacks and mitigations, and make coverage comparable across teams and tools.

Clear ownership, documented risk decisions, review before launch and regular reassessment, so security isn't left to whoever happens to build each feature.