What “air-gapped” really means for an AI deployment
Many vendors describe their AI as “private”. For a regulator or a bank, the question is narrower: is there any network path from the system to the outside world? In an air-gapped deployment the answer is no, and every part of the stack has to be designed around that.
Inside the perimeter
- Model weights, from a local registry
- Retrieval indexes and documents
- Prompts, answers and traces
- Audit logs, written once
- Identity, through your SAML or OIDC
How updates come in
What stays inside
- Model weights, served from a local registry.
- Retrieval indexes and the documents behind them.
- Prompts, answers and traces.
- Audit logs, written once and never edited.
- Identity, through your own SAML or OIDC provider.
How updates arrive
New model versions and fixes arrive as signed packages through your organization’s controlled transfer process. Nothing pulls from the internet. Each package is verified before it is installed, and the previous version stays available for rollback.
How we evaluate without the cloud
Every deployment ships with its own evaluation suite, built from your documents and your users’ questions. It runs inside the perimeter in staging, then again during a parallel run, before anything reaches production.
Questions to ask any vendor
- Which component needs an outbound connection, and why?
- Where are prompts and answers logged, and who can read them?
- How is a model update delivered and verified?
- Can the system run for a month with the network cable unplugged?