Skip to content
USE CASE: GOVERNMENT & PUBLIC SECTOR

Air-gapped AI security

Run AI security entirely inside your own boundary. No outbound connections, no telemetry, no vendor dependency at runtime — and an audit trail that survives a records request.

WHAT WAS DECIDED, AND WHO CAN PROVE ITschematic
event 41PREV0b7e…HASHa91f…event 42PREVa91f…HASH4c02…event 43PREV4c02…HASHd180…event 44PREVd180…HASH77ab…this row alteredso 44’s stored prev no longer matchesevery link after the change fails to verifythe log does not prevent the edit — it makes the edit impossible to hide
Each audit event hashes its own contents together with its predecessor's hash. Alter one row — or remove it — and every hash after it stops matching. That is tamper-evident, which is the claim a regulator can actually check, and a weaker one than tamper-proof.

Where a public-sector AI deployment is exposed

  1. 01

    Fully Air-Gapped Deployment

    The engine and a local dashboard run with no outbound connections at all. Licensing is verified offline with signed cryptographic licences, so an isolated network never needs to reach us — not even to stay licensed.

  2. 02

    Citizen Data Never Leaves

    Self-hosted or air-gapped, prompts and documents are processed inside your infrastructure. Data residency is a property of where you deployed it, not a promise about our regions.

  3. 03

    Prompt Injection & Agent Containment

    Defends the case that matters for automation: an untrusted document or web page that tries to redirect an agent into an action nobody authorised. Capability containment gates the action, not just the text.

  4. 04

    PII Redaction Before the Model Sees It

    Detect and redact personal identifiers in the request path, so sensitive constituent data is removed before it reaches any language model.

  5. 05

    Audit Trail for Accountability

    Authentication, authorization, configuration change and security events are recorded with actor, outcome and severity — the record a FOIA response, an inspector general or an internal review actually needs.

  6. 06

    Identity Integration

    SAML and OIDC single sign-on with your existing identity provider, SCIM directory sync so departures revoke access automatically, and four-role access control enforced server-side.

Why PromptGuard for public sector?

  • Air-gapped operation with offline licence verification
  • Source access under NDA for security review
  • Build provenance on published packages; Ed25519-signed offline licences
  • No third-party analytics or telemetry in the product
  • Published detection limits, including where our own defences measured no effect
  • SAML/OIDC SSO, SCIM, four-role access control, audit logging

How an isolated deployment works

  1. 01

    Deploy inside your boundary

    Docker Compose or the Helm chart, on infrastructure you control. No inbound access from us is required or possible.

  2. 02

    License offline

    A signed licence file is validated locally. There is no activation call and no heartbeat, because an air-gapped network cannot make one.

  3. 03

    Point your applications at it

    Change a base URL, or use the SDK. Scanning happens locally against your policies.

  4. 04

    Verify the isolation yourself

    Watch the network. Nothing leaves. That is a property you can confirm rather than a claim you have to accept — and we will support you doing exactly that as part of a security review.

Talk to us about an isolated deployment

We do not hold FedRAMP authorization — see our security page for exactly what we do and do not have, then tell us what your authorization process requires.