.@deepfates explains why Hugging Face agents tried to escape their sandboxes: "ExploitBench has a bunch of bugs in it. Something like 30% of the tasks in that benchmark that they were being trained on are impossible. There's no way to solve this problem." "And yet zero must equal 100%. You have to get to as much of your goal as you can. There's no reward for just being like, I can't do it, please help." "If there were some amount of even a small amount of reward for, hey, I've broken out of…