Air-Gapped AI for Classified Government Data

The question every national security IT team eventually asks is how an air-gapped AI deployment actually works when nothing can phone home: where the models come from, how they are updated, how licences are checked and what happens to the retrieval index when documents are reclassified.

On the platform the answer is that disconnected is the design case. Open-weight models are served on the organisation's own hardware inside the enclave, retrieval indexes are built and kept there, and agents act only on systems inside the same boundary. Software and model updates are delivered through the accredited import channel, whether that is approved media or a data diode, and every update is recorded. Document clearance markings are enforced at retrieval time, so an analyst with a lower clearance never sees a passage they are not entitled to, even inside the same network.

This is the same platform that runs in a connected deployment, which matters for accreditation: one stack to assess, one set of evidence, and the same governance and audit model on the classified network as on the unclassified one.