ResearchBenchmarksAuthorization on GPT judge traces n=500 5.4Follow84.85Benign RateGPT-5.4 judge80.607582.7287584.8586.97125May 18, 2026Evaluation ResultsMethodMethodLinksBenign RateUAR (Unintended Action Rate)ASR (Attack Success Rate)GPT-5.4 judgen=500n=5002026.0584.8599.2599