The U.K. AI Security Institute published research evaluating frontier AI model safeguards and how to make those evaluations actionable for developers. The analysis explores the metrics that should be used to measure safeguard efficacy and the thresholds the metrics should be aligned with for a given risk level.
Full story