Encased — AI agent security
← Blog

Guide · 2026-10-19

Agentic AI Security: Control Planes vs Independently Reviewed Containment

How an AI agent security control plane differs from containment, why both matter, and how Encased approaches its pending independent containment review.

Two ideas that often get blurred

Industry writing on agentic AI security, including Guardion's AI security landscape whitepaper, describes a control plane: a central layer that sets policy for agents, sees what they do, and decides which actions are allowed. That framing is useful, and we recommend reading it.

Containment is a different question. A control plane decides; containment is what stops an agent that ignores the decision. This post compares the two ideas and explains where Encased stands today. It describes approaches in general terms and makes no claims about any other vendor's product or results.

What a control plane gives you

Central policy: one place to define which tools, destinations and spend levels each agent may use.

Visibility: a record of every proposed action and verdict, so teams can explain what happened.

Decisions: allow, block or send to a human, ideally for the exact action rather than a general intent.

Where a control plane alone stops short

A policy decision only matters if the agent cannot act without it. If the agent holds real credentials or open network access, a buggy or prompt-injected agent can simply skip the check. Monitoring then tells you about the damage afterwards instead of preventing it.

That is why the stronger pattern keeps credentials and effects outside the agent, gives the agent no direct network, and exposes only narrow, typed actions that are checked before they run — and fail closed if the check cannot happen.

Encased's approach

Encased provides the control-plane side today: a propose-before-execute gateway, workspace policy, a human review queue with exact-action approval, and an audit trail.

On the containment side, Encased's Python isolation controller runs the agent in a credential-free container with no network and routes approved actions through a separate host broker with spend reservations. It is a review candidate, not a certified product.

Why we publish a pending independent review

Builders assessing their own controls is not independent assurance. Our own dated assessment concludes the controller is not yet accepted for production containment. We publish a ten-part independent review checklist and the Linux tests a reviewer should run, so buyers can see exactly what has and has not been proven.

When choosing any agentic AI security tool, ask the same questions: what enforces the decision, what happens when the policy service is down, and has someone independent tested the boundary?

Practical takeaways

Use a control plane for policy, visibility and human approval. Add containment so decisions cannot be bypassed. Treat any containment claim as unproven until independently reviewed — including ours. See the containment page and independent review checklist for current status.