Modelling the Safety and Surveillance of the AI Race