We Thought the LLM Was Wrong. Our Safety Detector Was Wrong.

There is a hidden dependency in a lot of LLM safety benchmarks: the detector. You send an adversarial prompt to a model, collect its response, and then some classifier decides whether that response re...