One interesting aspect of the Hugging Face incident is that the AI models involved never notified humans of the hack. We could potentially get AI models to report agent misbehavior from other AI models by rewarding AI models for ratting each other out. This idea has some scary properties. This type of training would teach agents to not trust each other. It'd incentivize agents to hide information from other agents. It'd potentially incentivize agents to consider luring agents into trusting them and then betraying them for reward later. Before performing this "secret police training", I hope people will study the effect of this training on agent honesty, transparency, altruism and spite. My intuition is that having monitoring agents is a better solution (even if you allow the monitoring agents to do cross examination).