Firevault, offline secure storage, physically disconnected from the internet
Pricing

Control BlueprintCP-08

Control AI Systems

An AI agent may need tools and data, but it should not inherit unlimited standing reach to production systems.

Blueprint
CP-08
Primary modules
  • Firebreak
  • Validate
  • Execute
  • Isolate
  • Relay
Supporting
  • Lock
  • Archive
  • Unlink
  • Transfer
Where it fits
Agent, model and tooling access with a physical hard stop
CP-08 Control AI Systems, as drawnIllustration
Zone 0AI runtime and agentsModels, orchestration and autonomous agents acting at machine speed
Boundary 1
  • Firebreak
  • Validate
  • Relay
Severed
Severed by default. Opens as a validated, time-bound window for a named purpose.
Zone 1Tooling and action layerAPIs, automation and the systems an agent is allowed to change
Boundary 2
  • Isolate
  • Execute
  • Lock
Scoped
Separated at hardware level. One Execute signal is the hard stop, enforced physically.
Zone 2Production data and systemsRecords, shares and operational systems the work depends on
Held offlineModel and training asset vaultWeights, curated training sets and the record of every window, held offline and off the network the agent works on.Offline

Out-of-band command pathManagement Ethernet, cellular SMS or an authenticated API, off the production network.

The sequenceStep through how the modules work together, as the Blueprint sets them out.

The pattern

Keep authority over what an AI system can reach.

What it does

CP-08 separates an AI runtime from the tools that can act and the production systems those tools can reach. Validate checks the permitted request, Relay limits the connection window and Firebreak provides a physical disconnect at the chosen boundary when the approved stop action is invoked.

Execute carries out the agreed boundary action even where stopping the chosen network path cannot be left to the AI application itself. It does not determine whether an output is correct, erase a model or reverse an action already taken. Keep human approval, monitoring and application safeguards around the deployment.

Where it fits

For agent runtimes, model services and AI tooling in finance, healthcare and defence. Start with the actions the system can perform, the data it may use and the production connections that must remain under human authority.

Grounded in

the EU AI Act Art. 14-15, NIST AI RMF MANAGE-2.3, ISO/IEC 42001 A.6 and IEC 62443-3-3 SR 5.1. Control can help support the controls they describe.

Control for ai systemsRead the AI kill switch briefing

Zone by zone

How CP-08 controls AI systems.

Distinguish the model runtime, its action tools and the production assets. A request at one stage is not unrestricted permission for the next.

  1. Zone 0

    AI runtime and agents

    Agents and model runtimes generate requests. Their presence does not grant a standing connection to every available tool.

  2. Boundary 1

    Severed

    Validate the requested authority and open only the tooling window required for the approved task.

    • FirebreakPhysical sever. Primary in CP-08
    • ValidateIntegrity check. Primary in CP-08
    • RelayTime-bound path. Primary in CP-08
  3. Zone 1

    Tooling and action layer

    Tools and action services can turn a model request into a real operation. Their onward reach needs a separately defined boundary.

  4. Boundary 2

    Scoped

    Keep production reach physically controlled, with Execute able to invoke the agreed stop and Unlink removing inherited dependencies.

    • IsolateZone boundary. Primary in CP-08
    • ExecuteTrigger, act, terminate. Primary in CP-08
    • LockNamed access. Supporting in CP-08
  5. Zone 2

    Production data and systems

    Production records and services remain governed assets, not an unrestricted extension of the AI runtime.

  6. Offline

    Model and training asset vault

    Keep selected model baselines, training assets and authoritative recovery material separately offline; validate them before reuse.

    Offline Secure Storage®

Bill of modules

Firebreak, Validate, Execute, Isolate and Relay lead. Lock, Archive, Unlink and Transfer support.

Five primary modules govern AI reach and the stop action. Supporting modules constrain identities, dependencies, artefacts and retained records.

Module Layer In CP-08 Where it sits What it does here
Firebreak Control the path Primary Boundary 1 Physically disconnects the selected AI-to-production path when the authorised stop action requires it.
Validate Protect the asset Primary Boundary 1 Checks whether an AI tooling request has the purpose and authority permitted for the task.
Execute Control the path Primary Boundary 2 Invokes the pre-agreed physical stop action at the AI boundary.
Isolate Control the path Primary Boundary 2 Separates runtime, tooling and production environments so model access does not imply general operational reach.
Relay Control the path Primary Boundary 1 Provides the approved AI tooling or data-access window rather than standing connectivity.
Lock Protect the asset Supporting Boundary 2 Restricts identities and roles around model, tooling and production access.
Archive Protect the asset Supporting Across the Blueprint Preserves the designated AI access and control records for investigation and review.
Unlink Protect the asset Supporting Across the Blueprint Removes persistent dependencies that could give an agent an unintended route around the boundary.
Transfer Protect the asset Supporting Across the Blueprint Governs model artefacts and data permitted to move between the protected environments.

The AI Kill Switch applies Firebreak® to the selected AI network path within CP-08. Firebreak is supplied by Firevault within Control; the stop action governs reach, not all possible AI behaviour.Explore Firebreak®

All nine modules

Versus the alternatives

A software stop and a disconnected path do different jobs.

Application safeguards remain essential. A physical boundary adds authority over the selected connection without claiming to solve every AI risk.

Question Detect and respond Manual shutdown CP-08
Request authority Application policy decides allowed AI actions An operator decides whether to connect the runtime Validate checks the approved purpose and authority
Access window Credentials can permit standing tool access Staff connect for the required work Relay makes the approved connection temporary
Stop action Depends on the application’s implemented response Staff remove the selected cable Execute invokes Firebreak at the agreed AI boundary
Reach after the stop Other permissions or routes need assessment Other connections may still exist Isolate and Unlink address the mapped environment and dependencies
Model artefacts Application controls govern stored material Disconnection alone does not preserve a baseline Transfer and separately protected assets support governed reuse
Human responsibility Operators still assess AI behaviour Cable control does not judge an output Humans retain final authority; the physical control limits reach

Questions

Questions about the AI boundary.

The physical action disconnects the selected network path. It does not necessarily stop local computation, delete the model or reverse actions already performed. Define exactly what reach the boundary controls.

An application stop is valuable, but the chosen network path can also be governed outside that application. CP-08 provides a separately authorised physical action at the defined boundary.

Yes. Validate checks the authority for the request and Relay governs the permitted window. The design distinguishes authorised tooling from unrestricted standing access.

A human authority or an approved external signal initiates the agreed action. Execute carries it out locally. The model should not be the sole authority over its own production reach.

No. Output quality, model behaviour and application safety need their own controls. This Blueprint governs selected connections, requests and artefacts, not the truth or safety of every response.

Keep selected authoritative baselines separately from the live runtime and govern transfers back into use. Verify provenance and suitability before restoring or retraining from those assets.

All questions

Disconnect to Protect®

Keep the final connection decision human.

Bring the tools and production services your AI can reach. We will discuss permitted windows and the exact boundary an authorised stop controls.