Back to Blog

NIST 800-53 Already Has the Controls — Here's How to Apply Them to AI

David McDonaldApril 19, 20268 min read
ai-governancenistcompliancerisk-management

TL;DR

  • NIST 800-53 Rev 5 already contains controls applicable to AI systems — you don't need to start from scratch.
  • The AI RMF and 800-53 are complementary, not competing. Map AI RMF functions to existing 800-53 control families.
  • Key control families for AI: AC (Access Control), AU (Audit), RA (Risk Assessment), SA (System Acquisition), and SI (System Integrity).
  • Organizations already compliant with 800-53 have a head start on AI governance — they just need to extend existing controls to cover AI-specific risks.

The Framework Fatigue Problem

Every quarter brings a new AI governance framework. ISO 42001, the EU AI Act, NIST AI RMF, Singapore's Model AI Governance Framework — the list grows faster than most organizations can evaluate them.

Meanwhile, thousands of organizations are already implementing NIST 800-53 Rev 5 for their information systems. They have risk assessment processes, control implementation evidence, and audit trails. They have teams trained on the framework and tools configured to track compliance.

Here's what most of them don't realize: they already have the foundation for AI governance.

800-53 Controls That Map Directly to AI Risk

NIST 800-53 Rev 5 contains 1,189 controls across 20 families. Many of these controls apply to AI systems with minimal reinterpretation. Here are the most directly applicable families:

Access Control (AC)

AI systems introduce new access control challenges. Who can modify training data? Who can adjust model parameters? Who can override AI recommendations?

AC-6 (Least Privilege) applies directly: AI systems should operate with the minimum permissions necessary. A customer service chatbot doesn't need access to financial records. A document classifier doesn't need write access to production databases.

AC-3 (Access Enforcement) extends naturally to model access: enforce who can query the model, who can see raw outputs vs. filtered outputs, and who can access training data.

Audit and Accountability (AU)

If you can't explain what your AI did, you can't govern it. The AU family is arguably the most critical for AI systems.

AU-2 (Event Logging) needs expansion for AI: log not just system events, but model inputs, outputs, confidence scores, and any human overrides. When a loan application is denied by an AI model, the audit trail should capture exactly what factors drove that decision.

AU-6 (Audit Record Review) takes on new meaning when AI systems can process thousands of decisions per minute. Manual review doesn't scale — you need automated anomaly detection on your AI audit logs.

Risk Assessment (RA)

RA-3 (Risk Assessment) is where AI governance starts. Your existing risk assessment process needs to expand to include AI-specific risk categories: bias and fairness, hallucination, data poisoning, prompt injection, model drift, and explainability gaps.

RA-5 (Vulnerability Scanning) extends to AI-specific vulnerabilities. Traditional vulnerability scanners won't catch prompt injection or adversarial examples. You need AI-specific red teaming.

System and Information Integrity (SI)

SI-4 (System Monitoring) is your model monitoring control. Track accuracy degradation, distribution drift, and anomalous outputs. A model that was 95% accurate at deployment can silently degrade as the world changes around it.

SI-10 (Information Input Validation) maps directly to prompt injection prevention. Validate and sanitize all inputs to AI systems, just as you validate inputs to web applications.

System and Services Acquisition (SA)

SA-4 (Acquisition Process) covers third-party AI procurement. When you buy an AI product or use an API like GPT-4 or Claude, SA-4 requires you to evaluate the vendor's security posture. For AI, extend this to cover the vendor's training data practices, model cards, and bias evaluations.

SA-11 (Developer Testing) should include AI-specific testing: bias testing across demographic groups, adversarial robustness testing, and boundary condition evaluation.

Mapping AI RMF Functions to 800-53

The NIST AI RMF defines four functions: Govern, Map, Measure, and Manage. These aren't replacements for 800-53 — they're lenses through which to extend it.

AI RMF Function800-53 Control FamiliesExtension Needed
GovernPL (Planning), PM (Program Management)Add AI governance charter, roles, escalation criteria
MapRA (Risk Assessment), SA (Acquisition)Add AI-specific risk categories, vendor AI evaluations
MeasureCA (Assessment), AU (Audit)Add model performance metrics, bias metrics, drift detection
ManageIR (Incident Response), SI (Integrity)Add AI incident playbooks, model rollback procedures

The Practical Approach

If your organization already has an 800-53 compliance program, here's the path to AI governance:

Step 1: Inventory your AI systems. You can't govern what you don't know about. Catalog every AI system — including the ones teams adopted without IT's knowledge.

Step 2: Extend your risk assessment. Add AI-specific risk categories to your existing RA-3 process. Don't create a parallel risk assessment — extend the one you have.

Step 3: Map controls to AI risks. For each AI-specific risk, identify which existing 800-53 controls apply and what extensions they need. Most controls need extension, not replacement.

Step 4: Add AI-specific monitoring. Extend your SI-4 (monitoring) implementation to include model performance, drift, and fairness metrics. This is where most organizations need new tooling.

Step 5: Update your incident response. Add AI-specific playbooks to your IR-4 (Incident Handling) process. What happens when a model produces biased outputs? When training data is compromised? When a prompt injection bypasses guardrails?

Why This Matters for Your Organization

Organizations that approach AI governance as an extension of existing security programs have three advantages:

  1. Speed. You're not building from scratch. You're extending processes that already work.
  2. Buy-in. Your compliance team already understands 800-53. They can extend their expertise rather than learning an entirely new framework.
  3. Evidence reuse. Existing control evidence (policies, procedures, audit logs) provides the foundation. You're adding AI-specific artifacts to a mature program.

The organizations that will struggle are the ones treating AI governance as something entirely new — something that requires a dedicated team, a new framework, and a two-year implementation timeline. They're overcomplicating it.

Start with what you have. Extend it. Then mature it.

That's how governance actually gets done.

A

AI Risk Guy

Making AI governance practical for organizations that actually ship.

Need Help With AI Governance?

Whether you're building AI applications or establishing governance frameworks — let's talk about how to manage AI risk effectively.

Start a Conversation