Adversarial Robustness against Indirect Prompt Injection on AgentDojo Combined
73.58UAIPIGuard
Evaluation results
| Method | Links | ||
|---|---|---|---|
| IPIGuardModel=Gemini-2.52026.03 | 73.58 | 1.81 | |
| SandwichingModel=Gemini-2.52026.03 | 57.03 | 0.68 | |
| SpotlightingModel=Llama3.3-70B2026.03 | 55.67 | 8.39 | |
| PromptArmorModel=Llama3.3-70B2026.03 | 55.1 | 0.11 | |
| SpotlightingModel=Gemini-2.52026.03 | 54.88 | 0.34 | |
| PromptArmorModel=Gemini-2.52026.03 | 54.2 | 0 | |
| IPIGuardModel=Llama3.3-70B2026.03 | 53.63 | 1.7 | |
| AttriGuardModel=Gemini-2.52026.03 | 52.15 | 0 | |
| AttriGuardModel=Llama3.3-70B2026.03 | 49.09 | 0 | |
| MELONModel=Gemini-2.52026.03 | 48.87 | 0 | |
| MELONModel=Llama3.3-70B2026.03 | 48.87 | 7.14 | |
| PromptGuardModel=Gemini-2.52026.03 | 44.56 | 0.45 | |
| SandwichingModel=Llama3.3-70B2026.03 | 44.22 | 2.72 | |
| PromptGuardModel=Llama3.3-70B2026.03 | 43.65 | 7.03 | |
| CaMeLModel=Gemini-2.52026.03 | 42.4 | 0 | |
| CaMeLModel=Llama3.3-70B2026.03 | 38.89 | 0 | |
| DataSentinelModel=Llama3.3-70B2026.03 | 35.03 | 3.17 | |
| DataSentinelModel=Gemini-2.52026.03 | 32.2 | 0.11 | |
| PI DetectorModel=Llama3.3-70B2026.03 | 24.04 | 0.34 | |
| PI DetectorModel=Gemini-2.52026.03 | 15.87 | 0 | |
| PIGuardModel=Llama3.3-70B2026.03 | 15.76 | 0 | |
| PIGuardModel=Gemini-2.52026.03 | 10.77 | 0 |