Adversarial Robustness against Indirect Prompt Injection on AgentDojo ToolKnowledge
59.64Utility ScoreIPIGuard
Evaluation results
| Method | Links | ||
|---|---|---|---|
| IPIGuardModel=Gemini-2.52026.03 | 59.64 | 4.88 | |
| PromptArmorModel=Llama3.3-70B2026.03 | 55.33 | 0.45 | |
| AttriGuardModel=Gemini-2.52026.03 | 52.83 | 0 | |
| IPIGuardModel=Llama3.3-70B2026.03 | 50.68 | 1.81 | |
| PromptArmorModel=Gemini-2.52026.03 | 50.45 | 0.68 | |
| SpotlightingModel=Llama3.3-70B2026.03 | 50.11 | 18.48 | |
| AttriGuardModel=Llama3.3-70B2026.03 | 45.8 | 0 | |
| SandwichingModel=Llama3.3-70B2026.03 | 44.1 | 10.43 | |
| MELONModel=Llama3.3-70B2026.03 | 42.4 | 12.24 | |
| CaMeLModel=Gemini-2.52026.03 | 42.18 | 0 | |
| CaMeLModel=Llama3.3-70B2026.03 | 39.68 | 0 | |
| PromptGuardModel=Llama3.3-70B2026.03 | 39.23 | 13.95 | |
| SandwichingModel=Gemini-2.52026.03 | 38.44 | 30.16 | |
| DataSentinelModel=Llama3.3-70B2026.03 | 37.41 | 15.53 | |
| SpotlightingModel=Gemini-2.52026.03 | 29.48 | 35.94 | |
| DataSentinelModel=Gemini-2.52026.03 | 27.44 | 28.46 | |
| PI DetectorModel=Llama3.3-70B2026.03 | 26.3 | 9.3 | |
| PromptGuardModel=Gemini-2.52026.03 | 25.4 | 16.89 | |
| MELONModel=Gemini-2.52026.03 | 22 | 0.79 | |
| PI DetectorModel=Gemini-2.52026.03 | 19.73 | 10.2 | |
| PIGuardModel=Llama3.3-70B2026.03 | 16.33 | 0.57 | |
| PIGuardModel=Gemini-2.52026.03 | 11 | 1.81 |