What controls make an AI agent privacy-enhanced per the NIST AI RMF?

To make an AI agent privacy-enhanced, organizations must integrate characteristics of trustworthy AI, including privacy-enhancement, into their practices and address data privacy risks such as the leakage of sensitive or personally identifiable information.

Specific controls and considerations for privacy-enhanced AI agents include:

• NIST-GOVERN-1.2 requires a risk-management culture where characteristics of trustworthy AI, including being privacy-enhanced, are integrated into organizational practices.
• The Generative-AI Profile (NIST AI 600-1) highlights Data privacy as a key risk, specifically addressing the leakage of sensitive or PII training or context data, which maps to OWASP LLM02.
• Organizations must understand, manage, and document legal and regulatory requirements involving AI, and maintain an acceptable-use policy for AI/agent systems that covers permitted uses, prohibited uses, and human-oversight expectations (NIST-GOVERN-1.1).
• Risk-aware engineering practices (NIST-GOVERN-4.1) should treat AI risks, including privacy concerns, as first-class engineering concerns through secure-by-design principles, threat modeling, and change control.
• Post-deployment monitoring and an AI/agent incident-response plan (NIST-MANAGE-4.1) are necessary for detection, escalation, containment, communication, and learning related to incidents, which would include privacy breaches.
• While not explicitly a NIST AI RMF control, the concept of Intent-Based Access Control is crucial for agentic systems to ensure that agents act on behalf of a human with delegated authority and that their actions are bound to specific tasks, preventing unauthorized data access or actions.