1 post
An AI sandbox escape is the wrong thing to fear. The real AI safety risk is a system acting on a flattened, ungrounded model of a sensitive region.