Control AI Systems
The AI Control Blueprint: bound what an AI system can reach, and keep a stop that the model cannot argue with.
- CP-08
- Path and asset together
What this Blueprint does.
Agent, model and future superintelligence access with a physical AI kill switch
The AI Control Blueprint: bound what an AI system can reach, and keep a stop that the model cannot argue with.
Agent, model and future superintelligence access with a physical AI kill switch
Financial services, Defence, Critical infrastructure, Healthcare, Public sector, Technology
How CP-08 controls AI systems.
A FIRE+VAULT pattern. The AI runtime holds no standing route to production data or tooling. Reach exists only as a validated, time-bound window, and a single Execute signal severs it at the physical layer regardless of how the model or its orchestration behaves.
AI runtime and agents
Models, orchestration
Models, orchestration and autonomous agents acting at machine speed
Firebreak
Validate
RelaySevered by default. Opens as a validated, time-bound window for a named purpose.
Tooling and action layer
APIs, automation
APIs, automation and the systems an agent is allowed to change
Isolate
Execute
LockSeparated at hardware level. One Execute signal is the hard stop, enforced physically.
Production data and systems
Records, shares
Records, shares and operational systems the work depends on
Crown jewels · detail callout
Model and training asset vault
Weights, curated training sets and the record of every window held offline, outside anything an agent or its operator can reach.
Modules & symbols
FirebreakPhysical severValidateIntegrity checkRelayTime-bound pathIsolateZone boundaryExecuteApproved actionLockNamed accessHow the CP-08 pattern composes.
- 1
FirebreakFIRE
Physically breaks the connection path so the attack cannot continue.
- 2
ValidateVAULT
Checks the request, command or approval before anything proceeds.
- 3
ExecuteFIRE
Fires the control action the moment a signal demands it.
- 4
IsolateFIRE
Separates the affected environment into a controlled zone.
- 5
RelayFIRE
Opens a temporary, time-bound window for an approved purpose.
The Blueprint in full.
Firebreak holds the route between the AI system and production systems severed by default, so an agent inherits no standing path. Validate checks each request against identity, purpose and authority before anything opens. Relay grants the reach as a time-bound window rather than a permanent credential. Isolate keeps the model runtime, the data it processes and the tooling it can act on in separate zones. Execute is the hard stop: one signal severs the path at the physical layer, whether or not the software responds. Lock, Archive, Unlink and Transfer hold the models and training material behind identity controls, preserve the record of every window, remove always-on dependencies and govern what data moves in or out.
A local action, agreed before it is needed.
As AI agents gain authority to change systems, the agent itself becomes a high-value target. CP-08 keeps an independent physical way to disconnect an AI system, including a security or orchestration agent, from the systems it controls. AI can advise and act within limits; people retain final authority.
Where CP-08 applies.
Build control around your environment
Talk to our team about composing this Blueprint for your estate.
From £360 a month including VAT. 36-month commitment. First payment at checkout.