If you can align an agent towards "good" then it's reasonable to think you can align it towards "evil"