DeepMind’s approach to AGI safety and security splits threats into four categories. One solution could be a “monitor” AI.