DayNews.ai

AI safety tests may measure capability, not safety

A study finds popular agent-safety benchmarks often measure raw capability rather than genuine safe behavior.