2 min read

AI Governance Policy for Defense Contractors

A three-tier AI governance model for defense contractors: prohibited, restricted, and permitted AI tools, mapped to the CMMC Level 2 controls that AI usage touches.
AI Governance Policy for Defense Contractors
Photo by Markus Spiske / Unsplash

Tier 1 - Prohibited

AI tools that process, store, or transmit CUI/FCI. No exceptions without ATO.

  • Any AI tool that sends CUI to external servers
  • Consumer-grade AI assistants (ChatGPT, Claude) for CUI-related work
  • AI-powered translation services processing controlled documents
  • Cloud-based OCR/transcription on CUI documents

Tier 2 - Restricted

AI tools that operate in CUI environments but do not directly process CUI. Requires approval and monitoring.

  • AI-powered endpoint security tools in CUI enclaves
  • Code assistants used on projects adjacent to CUI systems
  • AI-driven SIEM/log analysis in GCC High environments
  • Automated vulnerability scanning with AI prioritization

Tier 3 - Permitted

AI tools operating entirely outside CUI boundaries on non-sensitive work.

  • AI writing assistants for marketing content
  • Code generation for non-CUI projects
  • AI-powered scheduling, HR, or finance tools
  • Training and education AI tools
ControlNameWhy AI usage touches it
AC.L2-3.1.1Authorized Access ControlGoverns who can access AI tools in CUI environments
AC.L2-3.1.2Transaction & Function ControlLimits what AI tools can do with controlled data
AT.L2-3.2.1Role-Based TrainingRequires AI-specific security training for CUI personnel
AU.L2-3.3.1System AuditingAudit logs must capture AI tool interactions in scope
CM.L2-3.4.1System BaselinesApproved AI tools become part of the system baseline
CM.L2-3.4.2Security Config EnforcementAI tool configurations must meet security requirements
MP.L2-3.8.1Media ProtectionAI-generated outputs containing CUI require marking and protection
SC.L2-3.13.1Boundary ProtectionNetwork boundaries must account for AI tool data flows
SI.L2-3.14.1Flaw RemediationAI tool vulnerabilities must be tracked and patched