The International Traffic in Arms Regulations turn 50 next year. When ITAR was promulgated in 1976, the regulated 'technical data' it governed was paper drawings in vaults, microfiche in archives, and conversations in SCIFs. The boundary between a controlled drawing and an unauthorized export was physical: a fence, a clearance, a transmittal letter. In 2026, that boundary has become functionally meaningless. Technical data lives in cloud buckets sprayed across global availability zones, in vector databases backing internal AI assistants, and in the weights of fine-tuned foundation models. ITAR's compliance regime, designed for an analog world, is being asked to govern a world it was never written for.

The Directorate of Defense Trade Controls has spent the past 18 months issuing a steady stream of advisory opinions, FAQ updates, and draft rules attempting to apply the existing statute to the new technological surface. Defense contractors who continue to treat ITAR as a checklist administered by a regulatory affairs function — rather than as a daily operational discipline embedded in DevOps, MLOps, and HR — are failing audits at record rates.

What Changed in 2025–2026

Three distinct DDTC actions reshape the compliance perimeter in 2026:

1. The Cloud Encryption Carve-Out Got Narrower

The 2020 'cloud carve-out' under §120.54 declared that end-to-end encrypted technical data, properly key-managed and not stored in §126.1 destinations, was not an export when transiting cloud infrastructure (22 CFR §120.54, 2024). The 2025 DDTC clarification narrows that carve-out in three important ways: BYOK key management must be 'demonstrably' under U.S. control; transit through Chinese-owned CDN edge nodes is now flagged; and SaaS providers acting as 'processors' of plaintext technical data are explicitly within scope (DDTC FAQ, 2025).

2. AI Model Weights Are Technical Data

A November 2025 DDTC advisory opinion confirmed that the weights of a model fine-tuned on USML-controlled technical data are themselves technical data under §120.10 (DDTC advisory, 2025). The implication is sweeping: every defense contractor running internal RAG or fine-tuning pipelines on controlled drawings has potentially produced a new ITAR-controlled artifact whose disclosure is governed exactly as the source documents were.

3. Foreign-National Engineer Access Tightened

The 'deemed export' doctrine — that disclosure to a foreign national in the U.S. is an export to that person's country — has been on the books since 1995. What changed in 2025 is enforcement reach: DDTC has begun routinely subpoenaing single-sign-on logs, repository access controls, and chat-platform mentions during audits (Justice Department, 2025). Companies cannot fall back on 'we briefed them not to look.'

Why GovCloud Alone Is Not Enough

AWS GovCloud, Azure Government, and Google Cloud Assured Workloads remain the foundation of any defensible cloud ITAR posture. But the providers themselves are clear that GovCloud is necessary, not sufficient. The customer is responsible for:

• Identity boundaries: U.S.-person-only access via authoritative HR-of-record integration. Federated identity from a non-GovCloud directory creates ambiguous attribution.
• Build pipeline integrity: CI/CD runners, container registries, and dependency caches must all live within the controlled enclave. A single PyPI mirror outside the boundary can re-export.
• Telemetry routing: observability data (metrics, traces, logs) often default to commercial regions. Sending stack traces of controlled software to a commercial Datadog tenant is a deemed export.
• Backup and DR: snapshot replication targets must remain in-region. A misconfigured cross-region replication policy ships your data to a non-controlled environment.

The AI Compliance Stack

Building defensible AI on controlled technical data demands a layered architecture that did not exist in any vendor's catalog 18 months ago. Mid-tier primes that built MVPs on commercial OpenAI tenants in 2024 are now scrambling to reproduce them on GovCloud Bedrock, Azure Government OpenAI, or fully self-hosted open-weight models.

Foundation Model Choice

The current shortlist for ITAR-compliant LLM access includes Azure Government OpenAI (with FedRAMP High and IL5 ATOs), AWS Bedrock in GovCloud (Anthropic Claude, Meta Llama, Amazon Titan), and self-hosted open-weight models (Llama 3.1, Mistral, Qwen) running on FedRAMP-authorized infrastructure. Each option has different export control implications, especially for fine-tuning.

Fine-Tuning and RAG

Retrieval-Augmented Generation that returns USML-controlled text snippets to a foreign-national engineer is a deemed export at the moment of retrieval. Companies must implement document-level access control all the way through the embedding store. Most current vector databases — including Pinecone, Weaviate, and Milvus deployments — were not architected for this granularity. Defense contractors are increasingly building custom retrieval layers that enforce per-document ITAR jurisdiction before the model ever sees the chunk (NIST AI Risk Management Framework, 2024).

What Mid-Tier Primes Should Do Now

• Map the AI surface area: every internal RAG, every Copilot deployment, every fine-tune. Most companies discover they have three to five times more AI integrations than the compliance team is tracking.
• Treat model weights as artifacts: track them in CMMC-compliant artifact stores with the same rigor as source code and drawings. Hash, version, and access-control them.
• Brief and re-brief foreign-national engineers: ambient AI exposes new disclosure surfaces (Copilot suggestions, AI-generated code, retrieved RAG context). Standard 'do not look' briefings do not cover this.
• Run tabletop audits: DDTC is no longer auditing as a snapshot. Walk the entire data lifecycle — ingest, embed, store, retrieve, generate, log — and find the leaks before the audit does.

Deemed Exports and the AI Copilot Problem

The single hardest compliance surface for the modern defense contractor is the GitHub Copilot–style AI coding assistant. These tools take an active context (the file you are editing, the repository structure, related files) and send it to a foundation model that may or may not run in a controlled enclave. The traditional deemed export analysis assumed a discrete, intentional disclosure to a named foreign national. Copilot-style tools disclose continuously and ambiently to an opaque model whose data residency is the customer's responsibility.

GitHub's GovCloud Copilot offering, AWS CodeWhisperer in GovCloud, and various Azure Government AI Studio integrations have begun to close the gap, but coverage is uneven, and many defense engineering teams still default to commercial Copilot tenants for productivity reasons. A 2025 internal audit at a large defense prime — disclosed in a DCSA Industrial Security Letter — found that more than 4,000 engineers across the company were using commercial AI coding tools against ITAR-controlled repositories (DCSA, 2025). The cost of bringing those workflows into compliance was substantial; the cost of being out of compliance was higher.

Cloud Vendor Concentration Risk

AWS GovCloud and Azure Government dominate ITAR-cleared cloud workloads, with Google Cloud Assured Workloads and Oracle FedRAMP-High footprints filling a meaningful but smaller share. The concentration is not in itself a compliance issue, but it is an operational continuity issue that the Department has begun to treat as such. A 2025 GAO review noted that 'cloud-vendor concentration in the defense industrial base now functionally exceeds the concentration of any historical defense systems integrator,' raising operational continuity concerns that historic policy frameworks did not contemplate (Government Accountability Office, 2025).

The practical implication is that contractors must now design for graceful degradation in the event of a single-cloud-vendor outage or compliance disruption — a level of cross-cloud resilience that most mid-tier primes have not architected.

Foreign-National Engineer Access in 2026

Foreign-national engineer access controls have tightened in three concrete ways since 2024. First, DDTC has narrowed the conditions under which a permanent-resident engineer is treated as a U.S. person for technical data access — green-card holders from §126.1 countries now require additional review. Second, the cloud audit trails subpoenaed during DDTC investigations now routinely capture not only direct file access but also model-mediated retrieval, code completions, and chat history. Third, the U.S. Citizenship and Immigration Services has accelerated information-sharing with DDTC on visa-status changes that affect ITAR access.

Defense companies with significant foreign-national engineering populations — and many do, particularly in software engineering disciplines — now operate access control systems that must update within hours of a visa-status change, not days. The HR-IT-security integration required to do this reliably is significant; the consequences of failing to do it are program-suspending.

The Compliance Reckoning Has Begun

ITAR was written in an age of paper and vaults. It will not be substantially rewritten any time soon. What is being rewritten — at speed, and largely through enforcement actions and advisory opinions rather than formal rulemaking — is what compliance looks like inside the modern defense contractor. The mid-tier primes that succeed in the next three years will be the ones that treat ITAR as a software design constraint, not a stack of binders. The ones that don't will be the ones whose programs get suspended, whose engineers can no longer touch a drawing, and whose customers find another supplier.