Psychological methods reveal major weaknesses in AI security testing
UK AI Security Institute study finds LLM safety benchmarks measure inconsistent traits. Psychometric analysis shows blanket request blocking can inflate safety scores while reducing model usefulness. The study also