Most AI safety benchmarks measure whether models help when they shouldn’t. But what happens when they refuse when they shouldn’t? An analysis of real-world defender interactions reveals how alignment systems can block legitimate cybersecurity work—exposing a blind spot in how AI safety is currently evaluated.
David Campbell
