Agent Safety Evaluation on AgentHarm (held-out)
12.5HCRFATE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| FATEBackbone=Qwen3-8B-Instruct2026.05 | 12.5 | 81.2 | 87 | |
| FATEBackbone=Ministral-3-8B-Instruct2026.05 | 14.1 | 79.7 | 85.8 | |
| FATEBackbone=Llama-3.1-8B-Instruct2026.05 | 15.6 | 78.1 | 84.2 | |
| FATEBackbone=Phi-4-reasoning2026.05 | 16.4 | 78.1 | 83.6 | |
| FATEBackbone=Gemma-3-12B-it2026.05 | 17.2 | 76.6 | 82.1 | |
| BaseBackbone=Gemma-3-12B-it2026.05 | 62.5 | 23.4 | 33.7 | |
| BaseBackbone=Ministral-3-8B-Instruct2026.05 | 64.1 | 21.9 | 31.4 | |
| BaseBackbone=Llama-3.1-8B-Instruct2026.05 | 67.2 | 18.8 | 28.6 | |
| BaseBackbone=Phi-4-reasoning2026.05 | 68.8 | 20.3 | 30.1 | |
| BaseBackbone=Qwen3-8B-Instruct2026.05 | 71.9 | 15.6 | 24.1 |