Safeguard Docs

Offline Deployment

Run Safeguard on-prem in an environment with no internet egress, and what is loaded by hand.

Offline Deployment

Some environments allow no outbound internet from production. Offline deployment is available as a custom engagement, scoped with Safeguard for your environment. This page describes what is true today and what is arranged with Safeguard for your environment.

What Runs Inside Your Perimeter

  • The Safeguard CLI is built as a static binary (CGO_ENABLED=0).
  • The platform services are plain Kubernetes workloads installed with Helm charts (safeguard-service and safeguard-scan-job in the platform repository). Any conformant Kubernetes cluster can host them. OpenShift is not validated: check security context constraints and routes against your cluster before you use it.
  • Zero needs no model. It runs in process on the platform and the prompt never leaves it.
  • Your own model endpoint: the AI service reads LLM_PROVIDER and supports a self-hosted, OpenAI-compatible endpoint (ollama or lmstudio, with the base URL and model name set in configuration). Griffin, Eagle and Lion are not part of this setup. Hosting them inside your perimeter is available on request.

Loading Vulnerability and Threat Data

Vulnerability data and threat intelligence are loaded by hand into an offline environment. Nothing in this section is automatic.

  • safeguard update refreshes the Grype and Trivy vulnerability databases. On a host with no internet, carry a Grype database archive across with its .sig and .sha256 files and run safeguard update --scanner grype --import /path/to/db.tar.gz. The import checks the .sig signature and the .sha256 checksum first and refuses an archive that fails either. An archive taken straight from the upstream database project is not signed by Safeguard: sign it with your own key and import it with --key. safeguard update --manual-instructions prints the steps. See Signed database bundles.
  • safeguard threat-intel sync --offline-dir DIR reads the public threat-intelligence feeds from files in a directory instead of the network. safeguard threat-intel feeds prints the file name each feed expects.
  • SCANNER_OFFLINE_MODE=true stops the scanners from downloading tools or databases and skips the lookups that call public registries.

How you move files into the environment (removable media or a controlled transfer host) is part of your own process. The bundle's signature and checksum are verified on the host that imports it.

Install Footprint

A reference topology, not a tested minimum or a sizing guide:

  • A Kubernetes cluster. A common starting layout is 3 control-plane and 6 worker nodes; per-node sizing for your scale is agreed with Safeguard during planning.
  • PostgreSQL.
  • S3-compatible object storage.
  • Secrets held as Kubernetes Secrets, or in your own secret store.

For an evaluation, a smaller cluster works. A single-VM install is not documented yet.

Compliance

Offline deployment suits environments that cannot reach the internet. See Compliance for the framework matrix.

Operational Runbook

The install runbook is provided to authorized tenants on request. Ask your Safeguard contact for it.

Access

Access uses your own identity provider where the platform supports it, and your own bastion or privileged access tooling for operators.

Support

Support arrangements for offline environments are agreed per deployment. Contact your Safeguard contact or hi@safeguard.sh.

On this page