“The typical AI developer plan for addressing loss-of-control risk involves using pragmatic safety measures and monitoring to buy time until automated safety research can find better solutions.”