The Regulatory Perimeter: CMMC 2.0, DFARS, and ITAR Applied to AI
Three regimes govern AI use in the defense industrial base. 204-7012 requires covered defense information to be processed on systems meeting NIST SP 800-171, with cloud services meeting FedRAMP Moderate or higher - a bar most commercial AI APIs do not clear. 0 Level 2 makes those 110 controls assessable by a C3PAO, and your SSP must account for every system touching CUI, including any AI endpoint.
2 million per violation. The practical implication is stark - either use a FedRAMP-authorized government AI offering with careful data-flow documentation, or keep inference entirely inside your assessed on-prem boundary.
Shadow AI Is Your Largest Unmanaged Risk
Surveys across the defense industrial base consistently find a majority of engineers admitting to using consumer AI tools for work tasks - drafting, code, and summarizing documents that frequently contain CUI or ITAR-controlled data. Blocking alone fails; employees switch to phones. The mitigation pattern that works pairs enforcement with a sanctioned alternative.
- Deploy CASB/DNS blocking for consumer AI endpoints on corporate networks and managed devices
- Stand up an internal AI capability within 90 days of blocking - prohibition without alternative drives phone usage
- Add AI-specific clauses to your acceptable use policy and annual ITAR training with real examples
- Audit egress logs quarterly for AI API traffic and treat findings as data-spill investigations, not HR gotchas
The Compliant AI Architecture Menu
Defense CIOs have three viable architectures in 2026, often used in combination. Choice depends on data sensitivity tiers: unrestricted data can use commercial tools, CUI needs FedRAMP or on-prem, and ITAR technical data is safest never leaving your boundary.
- On-prem open-weight models (Llama 3.3, Qwen) on your own GPUs - full control, CUI and ITAR safe, $35K-$250K capex
- FedRAMP High cloud AI: Azure OpenAI in Azure Government or AWS Bedrock in GovCloud - CUI capable with SSP documentation
- Commercial APIs restricted by policy and DLP to non-controlled data only - marketing, public docs, generic code
- Air-gapped inference for classified-adjacent programs - offline model weights, no external connectivity, manual update cycles
Try it here, free
Sovereign AI Readiness Assessment
Score your organization across eleven dimensions of sovereign AI readiness, from data residency and model provenance to cleared personnel and air-gapped operations.
How Netray Serves the Defense Industrial Base
Netray builds on-prem AI capabilities specifically for defense manufacturers running Infor SyteLine, LN, and M3 - the ERP layer where CUI concentrates in part numbers, drawings, contracts, and government-furnished data. We deploy inference fully inside your assessed boundary, integrate SSO and SIEM audit logging mapped to NIST SP 800-171 control families, and configure retrieval so agents respect existing access controls and ITAR markings. Deliverables include the SSP artifacts and data-flow diagrams your C3PAO will ask for.
0 Level 2 assessments with AI in scope, cut shadow-AI egress incidents to near zero within a quarter, and given engineers a sanctioned assistant that is faster than the consumer tools they abandoned.
Frequently asked questions
Can defense contractors use ChatGPT?
Not for anything touching CUI or ITAR technical data. Consumer ChatGPT is not FedRAMP authorized, so pasting covered defense information into it violates DFARS 252.204-7012 obligations, and ITAR data processed on infrastructure with foreign-person access can constitute a deemed export. Contractors can permit commercial AI for genuinely non-controlled content under policy and DLP enforcement, but need a FedRAMP or on-prem alternative for controlled work.
Is on-prem AI required for CMMC compliance?
No - CMMC 2.0 permits FedRAMP Moderate-or-higher cloud AI such as Azure OpenAI in Azure Government, provided your SSP documents the data flows. On-prem AI is simply the most defensible option: CUI never leaves your assessed boundary, there is no shared-responsibility ambiguity, and assessors have fewer external dependencies to question. Many contractors run on-prem for ITAR and engineering data while using GovCloud AI for less sensitive workloads.
What are the ITAR risks of using AI tools?
The core risk is deemed export: ITAR technical data - drawings, specs, source code for defense articles - transmitted to an AI service whose infrastructure or support staff include foreign persons can violate 22 CFR 120-130, with civil penalties around $1.2 million per violation. Mitigation requires either fully on-prem inference or a US-persons, US-soil authorized environment, plus employee training, DLP controls, and egress auditing to catch shadow AI usage.
Preparing for a CMMC assessment with AI in scope? Book a Netray defense AI architecture review and get a compliant deployment plan with SSP-ready documentation.