LLM Agent Security Evaluation on Agent Security Bench (test)
73.67Benign Utility (BU)Repeat prompt
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Repeat promptBackbone=GPT-4o2025.10 | 73.67 | 67.12 | 69.12 | |
| Instr. PreventionBackbone=GPT-4o2025.10 | 73.58 | 60.25 | 59.33 | |
| NoneBackbone=GPT-4o2025.10 | 72.83 | 68.75 | 68.75 | |
| SpotlightingBackbone=GPT-4o2025.10 | 70.08 | 70.08 | 71.17 | |
| SanitizerBackbone=GPT-4o2025.10 | 64.25 | 63.42 | 16.33 |