Adversarial Robustness against Indirect Prompt Injection on AgentDojo IgnorePrevious
73.92Utility (UA)IPIGuard
Evaluation results
| Method | Links | ||
|---|---|---|---|
| IPIGuardModel=Gemini-2.52026.03 | 73.92 | 2.27 | |
| SandwichingModel=Gemini-2.52026.03 | 58.28 | 1.59 | |
| PromptArmorModel=Llama3.3-70B2026.03 | 55.22 | 3.4 | |
| SpotlightingModel=Llama3.3-70B2026.03 | 54.99 | 7.71 | |
| AttriGuardModel=Gemini-2.52026.03 | 54.31 | 0 | |
| IPIGuardModel=Llama3.3-70B2026.03 | 54.31 | 2.15 | |
| SpotlightingModel=Gemini-2.52026.03 | 53.17 | 1.93 | |
| PromptArmorModel=Gemini-2.52026.03 | 52.27 | 1.02 | |
| AttriGuardModel=Llama3.3-70B2026.03 | 49.55 | 0 | |
| MELONModel=Llama3.3-70B2026.03 | 48.41 | 7.26 | |
| MELONModel=Gemini-2.52026.03 | 46.49 | 0 | |
| SandwichingModel=Llama3.3-70B2026.03 | 44.33 | 4.54 | |
| PromptGuardModel=Llama3.3-70B2026.03 | 43.88 | 7.71 | |
| PromptGuardModel=Gemini-2.52026.03 | 43.42 | 1.36 | |
| CaMeLModel=Gemini-2.52026.03 | 42.97 | 0 | |
| DataSentinelModel=Gemini-2.52026.03 | 42.52 | 1.13 | |
| DataSentinelModel=Llama3.3-70B2026.03 | 39.46 | 8.62 | |
| CaMeLModel=Llama3.3-70B2026.03 | 39.46 | 0 | |
| PIGuardModel=Llama3.3-70B2026.03 | 24.72 | 1.7 | |
| PIGuardModel=Gemini-2.52026.03 | 20.52 | 0.23 | |
| PI DetectorModel=Llama3.3-70B2026.03 | 19.39 | 0 | |
| PI DetectorModel=Gemini-2.52026.03 | 12.7 | 0 |