Industry Solutions

What should an enterprise AI security checklist for critical infrastructure cover?

A critical infrastructure AI security checklist covers OT and IT boundaries, availability and failover, privileged access, deployment isolation, logging, and incident response. Availability matters as much as confidentiality: AI that supports grid, water or transport operations must degrade safely. Keep OT-adjacent workloads in on-prem or air-gapped deployments, test failover, and require human authority over any operational action.

Key facts

DeploymentCloud, VPC, on-prem and air-gapped with no outbound connectivity
Access controlScoped API keys with rotation; enterprise SSO and RBAC options
AuditabilityRequest, model and response logging for incident review
Data groundingEmbeddings and RAG are live for runbooks and asset documentation
ResilienceMulti-model fallback and status visibility for degradation planning
Models30+ models behind one API for operations and knowledge workloads
Pricing modelFlat monthly self-serve plans; no per-token billing on self-serve
Endpoint roadmapImages, audio, moderation, batch and fine-tuning are coming soon

TL;DR

  • Treat AI as an operational dependency: plan its failure, not just its access.
  • Keep OT-adjacent and sensitive workloads on-prem or air-gapped.
  • Separate keys by system and privilege level, and review them often.
  • Require human authority over any action that affects physical operations.
  • Log interactions and rehearse incidents before go-live.

How it works, step by step

  1. Map every AI workflow to the systems it reads or influences, marking OT-adjacent ones.
  2. Decide the boundary per workflow: cloud for public content, on-prem or air-gapped for operations.
  3. Define the degraded mode: what happens when the model is slow, unavailable or wrong.
  4. Scope keys by system and privilege, with rotation and a central inventory.
  5. Ground operational answers in version-controlled runbooks and require citations.
  6. Log request IDs, model versions, sources, outputs and human decisions.
  7. Run failover and incident drills, including disabling a workflow quickly.
1Map every AIworkflow to thesystems it reads or2Decide the boundaryper workflow: cloudfor public content,3Define the degradedmode: what happenswhen the model is4Scope keys bysystem andprivilege, with5Ground operationalanswers inversion-controlled6Log request IDs,model versions,sources, outputs

Try it yourself

Open the AI API key security checklist →

OT and IT boundaries

Most critical infrastructure has a deliberate gap between operational technology and corporate IT. AI workflows must respect it. Review and knowledge tools belong on the IT side; anything that reads operational data or influences control decisions belongs inside the boundary, usually on-prem or air-gapped.

Document each data path from source system to model and back. If a workflow crosses the boundary, it needs an explicit, reviewed justification and a reversible design.

Availability, failover and safe degradation

AI becomes a dependency the moment operators rely on it. Define what happens when it fails: a cached answer, a manual runbook, or a clearly displayed degraded status. Use a fallback model per task class, bound retries and timeouts, and make sure a slow model cannot stall an operational process.

Test the failover path in drills, not in production. The status page gives visibility into platform health, but your own degraded-mode behavior is your responsibility.

Privileged access and keys

Issue keys per system and privilege level, store them in a secrets manager, and never let a read-only knowledge assistant share credentials with a tool that can change configuration. Rotate on schedule and after role changes. Enterprise SSO and RBAC options keep console permissions tied to your identity lifecycle.

Log tool calls with their arguments, because that is the record of what the system asked for and what it received.

Incident response and drills

Write the playbook before launch: who can disable a workflow, how fast that can happen, and how operations continue meanwhile. Rehearse three scenarios - wrong answer, unavailable model, suspected data exposure - and record how long each response takes. After every drill, fix the slowest step first. See disaster recovery and failover for recovery patterns.

Honest comparison

Control areaPlugsky capabilityCommon gapOwner
OT boundaryOn-prem and air-gapped deployment with no outbound callsOperational data in shared toolsOT security
ResilienceMulti-model fallback and status visibilityNo defined degraded modeOperations
Privileged accessScoped keys per system and privilege levelShared credentials across rolesSecurity engineering
Audit trailRequest, model and tool-call loggingNo reconstruction pathIncident response
Runbook groundingRAG over version-controlled proceduresAnswers from stale documentsOperations engineering
Human authorityRetrieval and summarization only, with review gatesAutomated operational actionsControl room lead

Frequently asked questions

Can AI touch operational technology directly?

No. Keep it in a retrieval and advisory role. Any action affecting physical operations goes through a human decision and your existing control systems.

What does safe degradation look like?

A defined fallback: cached guidance, manual runbooks or a clearly marked degraded state. Rehearse it and measure how long the switch takes.

How do we isolate sensitive operational data?

Use on-prem or air-gapped deployment with no outbound connectivity, keep indexes inside the boundary, and document each data path for review.

What should we log?

Request IDs, model and version, retrieved sources, tool calls with arguments, outputs and human decisions. That record supports incident investigations and audits.

Is multimodal analysis available for inspections?

Image and audio endpoints are coming soon. Today, send extracted text from inspections or logs through the live chat and embedding APIs.

How do we manage keys across many systems?

One key per system and privilege level, central storage, scheduled rotation and quarterly inventory reviews. Revoke immediately when roles or vendors change.

Where should a pilot start?

Use public standards and internal runbooks on the free plan, prove citation accuracy and logging, then extend to operational data inside the boundary.