AI Security Compliance: NIST AI RMF, ISO 42001, and the EU AI Act

If you ship AI, three frameworks keep coming up: the NIST AI Risk Management Framework, ISO/IEC 42001, and the EU AI Act. They sound bureaucratic, but underneath they ask similar, sensible questions. This post maps them so you can start with the 20% that covers most of the intent.

The three you'll actually be asked about

NIST AI RMF (US, voluntary). A risk-management playbook built around four functions — Govern, Map, Measure, Manage. It's not a checklist to certify against; it's a way to structure how you identify and reduce AI risk.

ISO/IEC 42001 (international, certifiable). An "AI management system" standard, structurally similar to ISO 27001. If customers want a certificate, this is the one. It emphasizes policies, roles, an AI inventory, and continual improvement.

EU AI Act (EU, law). Risk-tiered regulation. Most business software lands in "limited" or "minimal" risk with transparency obligations; "high-risk" uses (e.g. hiring, credit, critical infrastructure) carry heavy requirements. Penalties are significant, so classification matters.

What they all really want

Strip away the vocabulary and the overlap is large:

1. Know your AI inventory. What models/features exist, what data they touch, who owns them.
2. Assess risk per use case. Impact if it misbehaves; likelihood; affected people.
3. Put controls in place. Input/output safeguards, human oversight, access control, logging.
4. Measure and monitor. Evaluate before launch and watch in production.
5. Govern it. Clear ownership, policies, and a review cadence.

If you do those five things and can show evidence, you're most of the way toward any of the three.

A pragmatic starting point

You don't need a GRC team to begin. Start with a self-assessment to find your gaps, then close the highest-impact ones:

• Run the Compliance Self-Assessment to get a baseline score and a prioritized gap list across governance, data, and security controls.
• For the technical controls that show up in every framework — input validation, data protection, logging — wire in concrete safeguards:
  • Input/output screening with the Prompt Firewall
  • Secret/PII redaction (see Detecting Secrets and PII in LLM Prompts)
  • Monitoring with LLM Observability

Map controls once, satisfy many

The efficient move is to implement a control once and map it to each framework. For example, "we screen and log all model inputs for injection and sensitive data" simultaneously supports NIST's Measure/Manage, ISO 42001's operational controls, and the EU AI Act's transparency and record-keeping expectations.

Keep a lightweight matrix: control → evidence → which framework clause it satisfies. That single artifact turns audits from a scramble into a lookup.

Takeaways

• The frameworks differ in form but converge on the same five asks: inventory, risk assessment, controls, monitoring, governance.
• Start with a self-assessment, fix the biggest gaps, and implement technical controls once.
• Map each control to multiple frameworks so compliance compounds instead of duplicating work.

Compliance is a byproduct of doing AI security well — not a separate project. For the engineering side of that, the LLM Security Checklist and the OWASP Top 10 for LLMs are good next reads.