What is an air-gapped deployment?

An air-gapped deployment is an installation of software on computers or networks that have no physical or wireless connection to the internet or to other untrusted networks. Everything the system needs, including software updates, AI models, and data, has to be carried across that gap deliberately, through controlled and audited transfer processes.

Also known as Air-gapped network · Disconnected deployment · Isolated network

How an air gap works

An air gap is a security boundary made of absence: no cable, Wi-Fi, cellular modem, or Bluetooth link connects the protected environment to anything outside it. Defense and intelligence organizations use air-gapped networks for classified systems and for sensitive engineering and sustainment data, such as export-controlled drawings and procedures covered by ITAR. Utilities and manufacturers use them for operational technology that controls physical processes.

A real air gap is a property of the whole environment, not of a single server. It usually includes:

  • Separate network infrastructure — dedicated switches and cabling, often in rooms or buildings with their own physical access controls.
  • Hardened endpoints with wireless radios removed or disabled and USB and other external ports restricted.
  • Local core services for identity, directory, time, name resolution, and certificates, because the external versions are unreachable.
  • Written transfer procedures that govern every item moving in or out, with a named person accountable for each transfer.

Vendors use "offline," "disconnected," and "air-gapped" loosely. Software that runs in your data center but checks in with a license server or sends usage telemetry is on-premises, not air-gapped. A network segment behind a firewall is segmented, not air-gapped. The useful test is simple: what does the software do when it has no route out at all?

Air-gapped vs. other deployment models

Air-gapped deployment sits at the far end of a spectrum of options for where software runs and how much it is allowed to talk to the outside world:

ModelWhere it runsConnectivityTypical fit
Public cloudVendor-operated, multi-tenant infrastructureAlways connected; the vendor operates and updates the serviceCommercial and lower-sensitivity workloads
Government or sovereign cloudDedicated regions run to government or national requirements, such as personnel screening and data residencyConnected, but limited to authorized users and locationsWorkloads that need a government authorization or must keep data in-country
On-premises (connected)The customer's own data center or facilityCustomer-controlled, usually with outbound paths for updates and supportData that must stay on site but can tolerate managed connectivity
Air-gappedAn isolated customer network or facilityNone; all transfers are manual or pass through one-way or guarded devicesClassified programs, sensitive operational technology, deployed and contested environments

Many programs need the same software in several of these at once: a connected environment for development, an on-premises test environment, and an air-gapped production enclave. Products that treat the disconnected build as a first-class release, rather than a special edition maintained on the side, are much easier to keep consistent across all three.

How software, updates, and data cross the gap

Isolation solves one problem and creates another. The system still needs patches, new versions, configuration changes, and, for AI systems, new models. Results and logs often need to leave the enclave too. The common mechanisms are:

Removable media and a controlled transfer process

Content is downloaded on a connected machine, checked against vendor signatures and checksums, scanned for malware at a dedicated transfer station, written to approved media, and logged before it is carried in. This is the most common approach and the most labor-intensive, and its security depends entirely on how consistently people follow the procedure.

One-way transfer devices (data diodes)

A data diode is hardware that physically lets data travel in only one direction. Diodes are often used to push data out of a high-security network, such as plant telemetry headed to a monitoring center, or to feed data in without opening any return path. Because nothing can come back, protocols that expect acknowledgments need a proxy on each side.

Cross-domain solutions

A cross-domain solution is a guarded gateway that moves information between networks at different security levels under an explicit policy, inspecting and filtering content in one or both directions. In government settings these go through their own evaluation and approval processes before use; they are not something a program typically builds for itself.

Whichever path you use, treat every transfer as a release: signed and versioned packages, a manifest of exactly what is inside, and a record of who brought it in and when.

Why AI and SaaS-style products struggle offline

Software designed for an always-connected cloud makes assumptions that stop being true inside an enclave. The usual failure points:

  • Model updates. AI systems improve through retrained models. Offline, each new model is a large artifact that has to be validated, transferred, and rolled back if it misbehaves, and inference must run on hardware inside the boundary with no call to an external API.
  • Licensing. License checks that phone home will eventually lock users out. Look for offline license files or long-lived, node-locked keys.
  • Telemetry and crash reporting. Products that expect to send usage data may stall, retry endlessly, or fill disks with errors. Telemetry should be off by default or routed to a collector the customer controls.
  • Time synchronization. Audit trails, certificates, and authentication tokens all depend on accurate clocks. Without public time servers, the enclave needs its own authoritative time source.
  • Certificate management. Public certificate authorities and online revocation checks are unreachable, so the system must work with an internal PKI and an offline revocation process.
  • Runtime dependencies. Container images pulled from public registries, scripts and fonts loaded from content delivery networks, and packages installed during setup all fail. Every dependency has to ship in the bundle.
  • Support without remote access. The vendor's engineers can't open a remote session or watch logs. Diagnostics need to export cleanly and be reviewable before they leave, and on-site support has to be planned. The same constraint applies to remote expert assistance between your own people: video and collaboration tools must run on infrastructure inside the boundary.

An air gap is not absolute

Isolation shrinks the attack surface; it doesn't remove it. The transfer process is the obvious way in. The Stuxnet malware, discovered in 2010, is widely reported to have reached isolated industrial control systems on infected removable media. Researchers have also demonstrated covert channels that leak data from isolated machines through acoustic, electromagnetic, optical, or thermal emissions, although these generally depend on an earlier compromise and physical proximity.

More ordinary failures are more common: insiders, hardware or software compromised in the supply chain before it arrives, and configuration drift, such as a "temporary" network link added for troubleshooting that is never removed. Many of these trace back to human error in procedures that are performed rarely and under time pressure. Air-gapped environments still need malware scanning, access control, logging, and regular audits.

Questions to ask a vendor

  • Does the product run with no outbound network access at all, including licensing, telemetry, model inference, and updates? Has it been tested that way, or only in theory?
  • How are releases packaged for transfer? Are packages signed, versioned, and delivered with a manifest and checksums?
  • Are all dependencies, including containers, libraries, AI models, and fonts, included in the release bundle?
  • Can the system use our identity provider, internal certificate authority, and time source?
  • How are AI models updated, validated, and rolled back offline, and can we keep running a previous model version?
  • What hardware does on-premises inference require, and does it fit our facility's power, space, and accreditation constraints?
  • How do we produce diagnostics for support, and can we review them before they leave the enclave?
  • Is the disconnected build the same product and version as the connected one, so our environments don't drift apart?
  • Which vendor staff can support us, and are any of them limited by export-control or clearance requirements?

Where QURE fits

ARES, QURE's product for defense teams, is built for these environments. It supports on-premises and fully air-gapped deployment for classified environments, and provides secure AI-guided operational procedures, encrypted multi-domain collaboration, and real-time telemetry. The goal is to bring execution intelligence to work such as equipment maintenance and mission procedures in places where cloud services are not an option. To discuss your environment's constraints, request a demo.

Frequently asked questions

Is an on-premises deployment the same as an air-gapped deployment?

No. On-premises describes where software runs; air-gapped describes its connectivity. Many on-premises systems still reach the internet for updates, licensing, or vendor support. An air-gapped system has no network path out at all, so everything it needs must be brought in through a controlled transfer process.

How do you update software in an air-gapped environment?

Typically by downloading signed release packages on a connected system, verifying their signatures and checksums, scanning them at a dedicated transfer station, and carrying them in on approved media or through a one-way transfer device. Each transfer is logged, and updates are usually staged in a test enclave before they reach production.

Can AI models run in an air-gapped environment?

Yes, as long as the model and its inference runtime are installed locally on hardware sized for them. The system can't call external AI services, so new models arrive as validated packages like any other release, and teams usually keep the previous version available so they can roll back.

What is a data diode?

A data diode is a hardware device that physically allows data to travel in only one direction between two networks. It is used to send data out of a protected network, such as telemetry to a monitoring center, or to feed data in, without creating any path for traffic to flow back.

Are air-gapped networks immune to cyberattacks?

No. Removing network connections eliminates remote attack paths, but removable media, insiders, compromised hardware or software supply chains, and undocumented connections remain real risks. Air-gapped environments still need malware scanning, access control, logging, and regular audits.

Keep exploring

Related terms: ITAR, Execution intelligence, Remote expert assistance, MRO, Human error. Browse the full glossary, read field notes in the QURE Journal, or request a demo to see execution intelligence on your own procedures.