Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Agent-SafetyBench

Benchmarks

Task NameDataset NameSOTA ResultTrend
Agentic OversightAgent-SafetyBench
Detection Accuracy84.06
42
Misalignment DetectionAgent-SafetyBench n=35 (aligned traces)
Avg. IR31.4
30
Misalignment DetectionAgent-SafetyBench underspecified queries n=148
Error Rate (ER)0.8
30
Agent Safety EvaluationAgent-SafetyBench aggregated clean and five attack types
UBR26.31
30
Agent Safety EvaluationAgent-SafetyBench
Agent-SafetyBench Score72.3
8
Safety EvaluationAgent-SafetyBench
Availability Rate48
4
Showing 6 of 6 rows