Why Export Control Compliance for AI Agents Is a Runtime Problem
Export control regimes such as ITAR and EAR were built around predictable, human-mediated transfers of technical data: a document sent, a briefing given, a system accessed by a known individual. AI agents do not fit that model. An agent can retrieve, summarize, translate, and forward controlled technical data across dozens of tool calls in seconds, often in ways that were never explicitly authorized by a human reviewer at the time. Static policy documents describe what should happen; they cannot inspect a tool call before it executes. That gap is why compliance for agentic systems has to be enforced at runtime, at the moment each action occurs, rather than assumed from a written policy.
What Counts as Technical Data When an Agent Is Involved
ITAR and EAR definitions of technical data and technology extend to derivative materials, not only original source documents. When an agent summarizes a controlled engineering specification, translates it, or incorporates it into a generated response, that output can itself carry export control obligations. Enterprises need a clear, documented view of which datasets meet these definitions, and they need that classification to be evaluated continuously, since an agent may combine unrestricted inputs in a way that produces a controlled output.
Derivative Outputs Are Often Overlooked
Classification efforts frequently focus on source documents and miss the outputs an agent generates from them. A summary, a translated excerpt, or a synthesized answer drawing on controlled material can carry the same restrictions as the original, even though it was never separately reviewed.
The Encryption Exemption Model and What It Requires of Agent Systems
Some export control frameworks permit the transfer of technical data across borders when it remains properly encrypted and access to decryption is controlled. Applying that model to agent systems requires separating the ability to retrieve data from the ability to decrypt it. An agent, or the infrastructure running it, should not hold blanket decryption authority; decryption should depend on the same attribute checks (nationality, clearance, project assignment) applied to any other access request. This distinction between retrieval and decryption is a key architectural requirement, not a compliance formality.
Agent Identity and Least Privilege in Practice
Agents that operate under a shared service account or inherit a user's full permission set make it difficult to answer basic compliance questions: which agent accessed this data, on whose behalf, and under what authorization. Giving each agent a distinct, auditable identity, scoped to the minimum permissions its task requires, makes it possible to enforce least privilege at the point of action and to reconstruct exactly what happened after the fact. This is the same principle underlying least-privilege access control for human users, applied to non-human actors that can act far faster and across more systems than a person typically would.
Tradeoffs and Open Questions
Runtime enforcement introduces real tradeoffs. Attribute checks on every tool call add latency and complexity to agent workflows. Human-in-the-loop review, while necessary for borderline classification decisions, slows down the automation gains that motivated agent deployment in the first place. And because ITAR and EAR do not yet define agent behavior explicitly, enterprises are extrapolating from deemed export and encryption exemption rules written for a different technical context. Reasonable organizations may reach different conclusions about where to draw these lines, which is why documented decisions and defensible audit trails matter as much as the technical controls themselves.