CMMC and AI tools handling CUI

The program walkthrough is CMMC and AI with CUI. The control-ranking question is Which NIST SP 800-171 controls matter most when AI tools handle CUI?. This page is the tool-buyer brief: when an AI product (SaaS model API, fine-tune job, evaluation platform, prompt store) sees CUI, what the CMMC Program actually says.

The rule is about information systems, not model types

CMMC and NIST SP 800-171 apply to information systems. An AI system is in scope when it processes, stores or transmits controlled unclassified information (CUI) or protects such a system. We searched 32 CFR Part 170 and SP 800-171 Rev. 3 and found no AI-specific CMMC clause. There is no “CMMC for LLMs” annex on those pages.

DFARS 252.204-7021 requires a current CMMC status at the level the contracting officer specifies for systems used in performance that process, store or transmit FCI or CUI. DoD CIO says Phase I self-assessment requirements remain in place and that implementation is paused in Phase 1. Check that page before you treat a phase date as current.

Scoping the AI tool

32 CFR 170.19 is the scoping rule. Asset categories and generative-system notes are in CMMC Program scoping for generative systems that handle CUI. Practical questions for an AI vendor:

• Does any prompt, fine-tune corpus, retrieved document, evaluation set or log contain CUI?
• Which of the vendor’s systems process, store or transmit that CUI (training cluster, inference API, support console, backup)?
• Is the vendor’s environment in your CMMC assessment scope, or is CUI staying in your enclave?

A tool that never sees CUI is not in CMMC scope as an AI tool. Your own enclave still is, if it handles CUI.

What a CMMC platform can and cannot be

Secureframe Defense says it Protect CUI and meet CMMC requirements and that organizations can demonstrate compliance with CMMC 2.0. The CMMC page names a CMMC-compliant enclave to securely store and access CUI, Policies, SSP, and POA&Ms auto-generated and maintained based on your actual environment, and Defense Navigator for scoping. That is a CMMC/CUI compliance offering. It is not a listing for assuring machine-learning models, and it is not a mapping from CMMC to the DoD AI ethical principles.

Vanta names CMMC Level 1, 2, or 3 and says CMMC is mapped to NIST SP 800-171 Rev 2 (and 800-172 for Level 3). That is still an information-system assessment product.

Neither page makes CMMC a substitute for DoD Responsible AI or for DoD Directive 3000.09.

Level 2 self-assessment is still an assessment

32 CFR 170.16 is the Level 2 self-assessment and affirmation rule. A vendor’s “AI-generated SSP” is an artifact you still have to affirm. See CMMC Level 2 vs Level 3 for AI.

Questions

Is there a CMMC control unique to foundation models?

We searched 32 CFR Part 170 and SP 800-171 Rev. 3 and found no mention of artificial intelligence or machine learning. Scope is still whether the system processes, stores or transmits CUI.

Does buying a CMMC platform qualify the model?

No cited page says that. CMMC status attaches to information systems used in performance. Model T&E is a different conversation, on CDAO T&E and red-teaming.