Industry Solutions

What should an enterprise AI security checklist for defense cover?

A defense AI security checklist covers classification handling, air-gapped deployment, supply-chain assurance, personnel access, export controls, logging, and change control. The starting assumption is that nothing leaves the boundary: models, indexes and logs run inside the environment with no outbound connectivity. Every model and update enters through a controlled, reviewed process, and humans own every decision with operational effect.

Key facts

DeploymentOn-prem and air-gapped with no outbound connectivity
Access controlScoped API keys with rotation; enterprise SSO and RBAC options
AuditabilityRequest, model and response logging inside the boundary
Data groundingEmbeddings and RAG are live for doctrine, manuals and procedures
Models30+ models behind one OpenAI-compatible API
Change controlModel and index updates shipped as reviewed, signed artifacts on your schedule
Pricing modelFlat monthly self-serve plans; no per-token billing on self-serve
Endpoint roadmapAudio, images, moderation, batch and fine-tuning are coming soon

TL;DR

  • Assume no outbound connectivity: air-gap models, indexes and logs.
  • Control every update as a reviewed, signed artifact on your schedule.
  • Scope access by classification and role, with named accountability.
  • Keep humans in authority over any decision with operational effect.
  • Log interactions inside the boundary for audit and investigation.

How it works, step by step

  1. Define the classification levels and the boundary each level requires.
  2. Deploy the platform air-gapped so no prompt, document or log leaves the environment.
  3. Establish an update process: review, sign, test and promote model and index artifacts.
  4. Scope keys and console access by role, with rotation and a central inventory.
  5. Ground answers in doctrine and manuals with citations, and refuse beyond approved content.
  6. Record request IDs, model versions, sources, outputs and human decisions.
  7. Exercise incident and failover drills inside the boundary before go-live.
1Define theclassificationlevels and the2Deploy the platformair-gapped so noprompt, document or3Establish an updateprocess: review,sign, test and4Scope keys andconsole access byrole, with rotation5Ground answers indoctrine andmanuals with6Record request IDs,model versions,sources, outputs

Try it yourself

Open the sovereign AI readiness score →

Classification and boundary design

Start from the classification of the content, not the convenience of the tool. Each level maps to a boundary: public reference material may run on standard infrastructure, while controlled content runs inside an isolated environment with no external calls. Keep separate indexes per classification so retrieval cannot mix levels.

Document each data path and its approval. In a defense review, an undocumented path is a finding even when the destination is secure. Review the map quarterly.

Air-gapped operation and controlled updates

Air-gapped deployment runs model serving, retrieval and logging inside your network with no outbound connectivity. That creates a process question: how do improvements arrive? Treat model and index updates as artifacts that are reviewed, signed, tested against your evaluation set and promoted on your schedule. Never allow automatic updates across the boundary.

Keep a version registry so every answer can be tied to the model and index revision that produced it.

Supply chain and personnel assurance

Defense procurement scrutinizes the supply chain behind every component. Maintain an inventory of models, serving components and dependencies, with provenance and review records. Access follows the same discipline: named individuals, role-scoped keys, periodic revalidation, and immediate revocation on change. Enterprise SSO and RBAC options keep console permissions aligned with personnel systems.

Change control and human authority

Every change to a prompt, tool permission or model version goes through the same review as any operational software change. Keep evaluation evidence, known limitations and rollback plans on file. Regardless of automation level, a human commander owns any decision with operational effect; the assistant's role is retrieval, summarization and drafting from approved sources.

Honest comparison

Control areaPlugsky capabilityCommon gapOwner
BoundaryAir-gapped deployment with no outbound connectivityExternal API calls from isolated networksSecurity accreditation
UpdatesReviewed, signed artifacts promoted on your scheduleAutomatic vendor updatesConfiguration control
AccessScoped keys, rotation, enterprise SSO and RBAC optionsShared credentials by teamPersonnel security
Audit trailRequest, model and source logging inside the boundaryLogs stored externallyAudit
GroundingRAG with citations over approved doctrineModel answers from memoryOperations
AuthorityHuman command over operational decisionsAutomated action from AI outputCommand chain

Frequently asked questions

Can the platform run without any external connectivity?

Yes. Air-gapped deployment keeps serving, retrieval and logging inside your network with no outbound calls, and the OpenAI-compatible API works locally.

How do improvements reach an air-gapped environment?

As reviewed, signed artifacts promoted through your own process. Test each update against your evaluation set before it enters production.

Is multimodal input supported?

Image and audio endpoints are coming soon. Today, process signals and scans upstream and send extracted text through the live chat and embedding APIs.

How do we handle multiple classification levels?

Run separate environments and indexes per level, with access scoped by role. Keeping levels in one shared index is the most common boundary mistake.

What does the audit trail contain?

Request IDs, model and index versions, retrieved sources, outputs and human decisions, all stored inside the boundary under your retention policy.

Do humans stay in the loop?

Yes. The assistant retrieves, summarizes and drafts. Decisions with operational effect remain with accountable humans.

How do we start?

Run a contained pilot on unclassified reference material with a single team, validate citations, logging and failover, then expand classification by classification.