What is the NIST AI RMF and how do I apply it to AI agents?

The NIST AI Risk Management Framework (AI RMF 1.0) is a framework designed to help organizations manage risks associated with AI systems, structured around four functions: Govern, Map, Measure, and Manage. Applying it to AI agents involves establishing clear governance, identifying and assessing agent-specific risks, continuously monitoring their performance, and having mechanisms to respond to and mitigate identified risks.

To apply the NIST AI RMF to AI agents, consider the following controls:

• Govern: Establish an acceptable-use policy for AI/agent systems that covers permitted and prohibited uses, and human-oversight expectations (NIST-GOVERN-1.1). Assign a named risk owner or accountable executive for each deployed AI/agent system (NIST-GOVERN-2.1). Define policies for human oversight of AI, including override authority and the boundary of agent autonomy, which links to OWASP LLM06/LLM08 (excessive agency) (NIST-GOVERN-3.2). Treat AI risks as first-class engineering concerns through practices like secure-by-design and threat modeling (NIST-GOVERN-4.1).
• Map: Identify and inventory AI systems, their context, and associated risks. This includes understanding what AI agents are running in the environment, as the lack of such awareness is a governance gap.
• Measure: Analyze, track, and measure risks, including those specific to Generative AI, such as confabulation/hallucination, information security risks (e.g., prompt injection, data exfiltration, insecure tool use, which map to OWASP LLM01/LLM02/LLM06), data privacy concerns (e.g., leakage of sensitive data, mapping to OWASP LLM02), and dangerous content (NIST AI 600-1).
• Manage: Prioritize and respond to AI risks (NIST-MANAGE-1.3). Implement procedures to deactivate, roll back, or safely retire AI systems that exceed risk tolerances, effectively providing a kill-switch for agents (NIST-MANAGE-2.3). Establish post-deployment monitoring and an AI/agent incident-response plan for detection, escalation, containment, communication, and learning (NIST-MANAGE-4.1).