Air-Gap AI Governance: DLP Without Internet Access

Air-gap requirements are not theoretical. Defense contractors handling ITAR-controlled technical data cannot send prompts through cloud services. Government agencies processing CUI under NIST 800-171 need every byte to stay inside the authorization boundary. Healthcare systems with strict data residency mandates need AI governance that works without a single outbound connection.

These organizations still need AI. Engineers need code assistance. Analysts need summarization. Researchers need literature review tools. The question is not whether to adopt AI — it is how to govern AI traffic when the network has no path to the internet.

What Air-Gap Means for AI Governance

An air-gapped deployment is not just a firewall rule. It is a network architecture where the entire data plane — every component that touches prompt content, model responses, and detected entities — operates with zero external connectivity. No DNS resolution to public servers. No certificate validation against external CAs. No telemetry exfiltration.

For AI governance, this creates specific technical challenges that cloud-based solutions cannot address:

DLP inference must run locally. Cloud DLP services are off the table. Named entity recognition, pattern matching, and classification models need to execute on local hardware. This means shipping model weights, maintaining inference runtimes, and handling updates through offline transfer.

Policy enforcement must be self-contained. The policy engine cannot call home for rule updates, threshold changes, or new compliance framework mappings. The entire rule set — detection policies, routing rules, and action handlers — must be resident on the data plane.

Audit logging must be local-first. Compliance evidence cannot stream to a cloud aggregator. Logs stay inside the authorization boundary and integrate with whatever SIEM or log management infrastructure exists within the air-gapped network.

How Outpost Solves Air-Gap Governance

Arbitex Outpost is the on-premises data plane that runs the complete governance pipeline inside your network. Everything that touches AI traffic — the DLP pipeline, the policy engine, the audit subsystem, and the provider routing layer — runs as containers within your infrastructure.

Offline DLP Inference

Outpost runs DLP models using an optimized local inference engine. AI-powered detection models and all pattern recognizers execute locally on CPU or GPU, depending on your hardware allocation. No model weights are fetched from external endpoints. No inference calls leave the network.

The three-tier DLP pipeline — regex pattern matching, statistical NER, and transformer-based classification — runs identically to the cloud version. The same detection policies produce the same results, whether the data plane is in a cloud region or an air-gapped server room.

Certificate-Based Authentication

In an air-gapped environment, SAML and OIDC federation to external identity providers is not an option. Outpost uses mTLS with X.509 certificate chains for authentication. The full chain — leaf, intermediate, and root CA — is verified locally. Certificate rotation is managed through the Admin API without requiring external CA connectivity.

For organizations running their own PKI (common in defense and government environments), Outpost integrates with internal certificate authorities. The trust chain is entirely under your control.

Policy Sync Without Internet

Policies are authored in the Arbitex admin portal and exported as signed configuration bundles. These bundles are transferred to the air-gapped environment through approved data transfer mechanisms — whatever process your security team uses to move approved data across the boundary.

On the Outpost side, the configuration bundle is verified against a known signing key, validated for schema integrity, and applied atomically. If validation fails, the existing policy set remains active. There is no partial application and no fallback to a permissive default.

For ongoing operations, the Outpost operations guide documents the full policy lifecycle: authoring, export, transfer, validation, and application.

Heartbeat and Health Monitoring

Outpost emits OpenTelemetry metrics covering request rates, DLP inspection latency, policy evaluation times, and model provider health. In an air-gapped deployment, these metrics route to your internal monitoring stack — Grafana, Prometheus, or whatever OTLP-compatible collector runs inside the boundary.

The heartbeat system reports component health without requiring external connectivity. If inference latency degrades, if a model provider becomes unreachable, or if the policy engine encounters an error, the health system surfaces it through internal metrics and the Admin API health endpoint.

Use Cases: Who Needs Air-Gap AI Governance

Defense and Intelligence

ITAR and EAR regulations restrict the export of defense-related technical data. When engineers use AI tools for code generation, design review, or technical writing, every prompt potentially contains controlled technical data. An air-gapped Outpost ensures that DLP inspection catches controlled content before it reaches any model — and that no prompt data ever transits infrastructure outside the authorization boundary.

Classified Research

National laboratories and cleared research facilities operate at classification levels where data handling is governed by strict access controls. AI assistants are valuable for literature synthesis, data analysis, and documentation — but only if the governance layer operates within the same classification boundary as the data it inspects.

Healthcare with Data Residency Requirements

Some healthcare organizations operate under data residency mandates that go beyond HIPAA’s baseline requirements. State-level regulations, contractual obligations with healthcare networks, or internal policies may require that patient-adjacent data never leave a specific geographic or network boundary. Outpost provides the same DLP protections — detecting PHI patterns like MRNs, diagnostic codes, and patient identifiers — without any data leaving the facility.

Financial Services with On-Premises Mandates

Banks and financial institutions with on-premises compute requirements can run Outpost alongside their existing infrastructure. DLP policies detect PCI-DSS patterns, account numbers, and financial PII. Audit logs integrate with existing compliance reporting pipelines. The governance layer operates under the same change management and access control processes as the rest of the on-premises environment.

What You Trade and What You Gain

Air-gap deployments require more operational investment than cloud SaaS. You manage container infrastructure, allocate compute for DLP inference, handle certificate rotation, and manually transfer policy updates. Model updates — new recognizers, improved detection weights — arrive through the same offline transfer process as policy bundles.

What you gain is complete data sovereignty. No prompt content, no model response, no detected entity, and no audit record ever leaves your network. The governance layer is fully self-contained, fully auditable, and fully under your operational control.

For organizations where this is a requirement rather than a preference, Outpost is the architecture that makes AI adoption possible without compromising the security posture that defines how you operate.

Learn more about deployment options in the Outpost product page or the deployment architecture guide. For detailed operational procedures, see the Outpost operations guide.