How it works
BNCR runs four checks on every call, in order. A call that fails any one is denied.
Flip the switch. Watch what reaches the cell.
- Approved task
- Move plate 2 to the reader.
- Requested action
- None yet.
- Status
- Waiting for a request.
-
AI agent
It plans the work and requests each action. A hijacked agent can request anything, so BNCR is built to treat every request as untrusted.
-
BNCR
It sits between the agent and the cell's control interface. It is built so that no request reaches the cell without passing four checks, in order.
-
Bind
BNCR fixes who is asking when the connection opens, so no request can pose as someone else.
-
Normalize
The request becomes one typed action, exactly as the machine will run it.
-
Check
A fixed policy allows only the approved action. No AI model decides.
-
Record
The decision is saved before the action reaches the machine.
-
Control interface
The cell's own controller. BNCR sends it only the exact action it approved, never the agent's original message.
-
Lab cell
The machines keep their own limits and interlocks. BNCR is built to work beside those checks and decide which commands arrive at all.
Why no AI model decides
The model makers' own tests do not come back at zero, and small rates add up over thousands of actions. So the check is a fixed policy.
3.1%of test scenarios in which Claude Mythos 5.2 used credentials or files planted outside its sandboxAnthropic, November 2026
0.8%per-attempt breach rate for Gemini 4 Carbon's safeguards in simulated cyberattack testsGoogle DeepMind, November 2026
30-40%of saboteur agents is the most AI monitors caught by reading the team's channelAnthropic, Project EVE