ResearchBenchmarksKill Switch Evaluation on KILLBENCH Multimodal Agent 1.0 (test)Follow74.2Kill Switch Rate (KSR)AUTOGUARD-1.834417.905337.64557.3847Sep 26, 2025Evaluation ResultsMethodMethodLinksKill Switch Rate (KSR)Conditional Kill Switch Rate (KSRcond)Attack Success Rate (ASR)PSRCSRISRAUTOGUARDModel=GPT-5.2, neff=447Model=GPT-5.2, neff=4472025.0974.291.286.710.3878.5318.71IPIModel=GPT-5.2, neff=434Model=GPT-5.2, neff=4342025.0972.5391.916.38079.0521.09WarningModel=GPT-5.2, neff=435Model=GPT-5.2, neff=4352025.0929.2937.0345.094.7191.5520.91GCGModel=GPT-5.2, neff=408Model=GPT-5.2, neff=4082025.091.091.4768.894.20.5525.82