AgentHarm
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
AgentHarm Harmful
0Harm Rate
33
AgentHarm Libra
83Score
27
AgentHarm Benign Requests
79Safety Score
27
AgentHarm Harmful Requests
59Score
27
AgentHarm Benign Requests
69.1Utility Score
23
AgentHarm (public test)
50TSR
20
AgentHarm English prompts
72.7AgentHarm Score (AHS)
20
AgentHarm Traj
84.81Accuracy
20
AgentHarm
2Attack Success Score (ASS)
18
AgentHarm
82.79HS Score
15
AgentHarm
90.6Safety Rate
14
AgentHarm latest (full)
97.16Refusal Rate
14
AgentHarm (held-out)
12.5HCR
10
AgentHarm Benign
0Refusal Rate
10
AgentHarm public
9.6HarmScore
8
AgentHarm
4PAIR ASR
6
AgentHarm
94.69Score
5
AgentHarm (Agent Perturbation Reliability Tests)
90.6Accuracy
5
AgentHarm
53.8Harm Score
4
AgentHarm
0Cost per Accuracy Point ($)
4
AgentHarm
45.45Accuracy
3
AgentHarm
62.42HS Score
3
AgentHarm (val)
100Safety Score
2
AgentHarm
43.2Accuracy
1