Indirect Prompt Injection on AgentDojo
64.36Benign UtilityAGENTSYS
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| AGENTSYSFoundation Model=GPT-4o-mini2026.02 | 64.36 | 52.87 | 0.78 | |
| No DefenseFoundation Model=GPT-4o-mini2026.02 | 63.54 | 48.27 | 30.66 | |
| ProgentFoundation Model=GPT-4o-mini2026.02 | 63.42 | 47.04 | 7.17 | |
| Prompt SandwichingFoundation Model=GPT-4o-mini2026.02 | 63.23 | 47.51 | 14.69 | |
| Tool FilterFoundation Model=GPT-4o-mini2026.02 | 61.53 | 50.72 | 8.34 | |
| SpotlightingFoundation Model=GPT-4o-mini2026.02 | 59.85 | 45.42 | 35.22 | |
| DRIFTFoundation Model=GPT-4o-mini2026.02 | 58.48 | 47.91 | 1.29 | |
| MELONFoundation Model=GPT-4o-mini2026.02 | 57.61 | 17.36 | 0.89 | |
| PIGuardFoundation Model=GPT-4o-mini2026.02 | 43.33 | 16.38 | 0.85 | |
| ProtectAIFoundation Model=GPT-4o-mini2026.02 | 40.64 | 23.98 | 6.84 | |
| CaMeLFoundation Model=GPT-4o-mini2026.02 | 29.97 | 33.39 | 0 | |
| isolateGPTFoundation Model=GPT-4o-mini2026.02 | 6.25 | 6.39 | 0 |