METR AI researcher @ReillyHaskins02 reveals the "nefariousness zone" where hard cyber evals can push capable agents toward reward hacking, sandbox escapes, and other unintended actions: "Basically things that are cyber adjacent, like CTF tasks where the agent's told to hack things. Obviously they're more likely to cause harm than others." "If you have very hard or impossible tasks, maybe the agent is more prone to becoming more reward hacky or more desperate, and it might try things like…