NIST AI RMF HIPAA Security Rule Crosswalk
Compliance Guide

NIST AI RMF HIPAA Security Rule Crosswalk: Mapping Functions to Safeguards for AI Systems That Handle ePHI

The NIST AI RMF is a voluntary, risk-based structure. The HIPAA Security Rule is a set of required administrative, physical, and technical safeguards. The AI RMF functions (Govern, Map, Measure, Manage) can organize and support the risk analysis, risk management, and review work HIPAA expects, but they do not replace it.

This guide is not legal advice.

Treat the AI RMF as a supporting structure, not a substitute

Technical safeguards such as access control and audit controls must also be enforced at runtime, because policy documents alone do not control what an AI agent does with ePHI. The AI RMF helps organize the work; the Security Rule defines the safeguards that work must address.

How the two frameworks relate
ElementDescription
AI RMFVoluntary, risk-based. Govern, Map, Measure, Manage.
HIPAA Security RulePrescriptive safeguards under 45 CFR 164.308, 164.310, and 164.312.
Runtime enforcementAgent identity, least privilege, tool-call policy, and audit logging produce the evidence.

Function-level crosswalk

The four AI RMF functions (Govern, Map, Measure, Manage) can organize and support the risk analysis, risk management, and review work that HIPAA expects. Record each mapping as an AI RMF activity, a HIPAA standard, an evidence artifact, and a status.

Fields in a mapping record
FieldWhat to record
AI RMF activityThe Govern, Map, Measure, or Manage activity being mapped.
HIPAA standardThe Security Rule safeguard it supports (45 CFR 164.308, 164.310, or 164.312).
Evidence artifactThe artifact that demonstrates the control, such as defined audit log fields reviewed on a set cadence.
StatusSupports, partially supports, or gap.

Where the frameworks do not align one to one

The AI RMF is voluntary and risk-based, while the Security Rule prescribes safeguards. Mappings will therefore be partial in places, and some AI-specific risks are not named in the Security Rule at all. Document those risks with owners and treatment decisions rather than forcing them into a HIPAA standard.

The same gap appears between documentation and enforcement. Administrative controls can be documented, but technical safeguards cannot be satisfied by documentation alone.

Which mapped controls need runtime enforcement

Administrative controls can be documented. Technical safeguards cannot be satisfied by documentation alone, because an autonomous agent acts at machine speed and makes access decisions during execution. Four controls tend to carry most of the technical weight for agents that handle ePHI.

Agent identity
Gives each agent a distinct, attributable identity, so access control and audit records point to a specific actor rather than a shared service account.
Least-privilege permissions
Limit each agent to the data and tools its task requires, and are checked when access is attempted.
Tool-call policy enforcement
Evaluates each call against policy before it executes, with approval workflows for higher-risk actions.
Audit logging
Records each request, decision, and outcome in a form that can support information system activity review.

These controls serve both frameworks. They give Manage and Measure activities something concrete to test, and they give HIPAA access control and audit control requirements verifiable evidence. Runtime governance platforms, including Trussed AI, address these areas through agent identity, permissions, policy enforcement, and audit logging. Whatever tooling you choose, confirm that it enforces decisions at runtime and records them, and that your own risk analysis covers it.

Building the control-mapping and evidence approach

  • Inventory every AI agent and tool that creates, receives, maintains, or transmits ePHI, and confirm each is in the current risk analysis.
  • Record each mapping as AI RMF activity, HIPAA standard, evidence artifact, and status (supports, partially supports, gap).
  • Mark each control as policy-only or runtime-enforced, and treat technical safeguards as incomplete without enforcement.
  • Define which audit log fields and review cadence will serve as activity review evidence.
  • Document AI-specific risks the Security Rule does not name, with owners and treatment decisions.
  • Verify all mappings against the primary framework texts and have counsel review them.

Regulatory change and limits of this guide

Mappings depend on the current text of each framework, so verify them against the primary sources before relying on them.