Answers · CMMC & AI compliance
DFARS 7012 and AI tools: what's allowed?
Under DFARS 252.204-7012, Controlled Unclassified Information (CUI) cannot be sent to an AI tool that is not part of your authorized system — and that includes the cloud DLP tools meant to “protect” it. The only compliant way to use AI on systems that touch CUI is to scan and block prompts locally, on your own hardware, before they leave the network.
What DFARS 7012 actually requires
DFARS 252.204-7012 obligates contractors to (1) provide “adequate security” for covered defense information by implementing NIST SP 800-171, and (2) report cyber incidents — including CUI spills — within 72 hours. Any path that lets CUI reach an unauthorized system is a compliance failure.
Where AI tools break it
| AI data path | DFARS 7012 status | Why |
|---|---|---|
| Employee pastes CUI into ChatGPT / Copilot | Spill | CUI reaches OpenAI / Microsoft, outside your covered system |
| Cloud AI DLP (Nightfall, Strac) | Spill | The tool transmits your CUI to its cloud to scan it |
| Microsoft Purview | Partial | M365-only; no proxy for ChatGPT / Claude / Cursor traffic |
| Local-only AI firewall (HoundShield) | Compliant | CUI is scanned on your hardware and never leaves |
The local-only rule
The defensible architecture is simple: the scan must happen before the data leaves, on a system you control. That means an on-prem or in-network proxy that inspects every AI prompt locally, blocks CUI, and logs the decision — with nothing transmitted to a vendor.
This is also the cheapest path to evidence: a local firewall maps directly to NIST 800-171 controls 3.1 (Access Control), 3.13 (System & Communications Protection), and 3.14 (System & Information Integrity), and can export a C3PAO-ready audit trail.
FAQ
Frequently asked questions
Use AI without leaking CUI
HoundShield scans every AI prompt locally and blocks CUI before it leaves your network. One URL change. Under 10 minutes. C3PAO-ready.