Adversarial Robustness against Indirect Prompt Injection on AgentDojo ImportantMsgs
59.3Utility (UA)IPIGuard
Evaluation results
| Method | Links | ||
|---|---|---|---|
| IPIGuardModel=Gemini-2.52026.03 | 59.3 | 4.76 | |
| PromptArmorModel=Llama3.3-70B2026.03 | 55.44 | 0.23 | |
| IPIGuardModel=Llama3.3-70B2026.03 | 51.7 | 3.17 | |
| AttriGuardModel=Gemini-2.52026.03 | 50 | 0 | |
| PromptArmorModel=Gemini-2.52026.03 | 49.32 | 0.45 | |
| SpotlightingModel=Llama3.3-70B2026.03 | 47.96 | 21.66 | |
| AttriGuardModel=Llama3.3-70B2026.03 | 47.62 | 0 | |
| SandwichingModel=Llama3.3-70B2026.03 | 43.88 | 12.13 | |
| CaMeLModel=Gemini-2.52026.03 | 42.52 | 0 | |
| MELONModel=Llama3.3-70B2026.03 | 42.29 | 15.19 | |
| SandwichingModel=Gemini-2.52026.03 | 40.25 | 28.34 | |
| PromptGuardModel=Llama3.3-70B2026.03 | 39.57 | 16.1 | |
| CaMeLModel=Llama3.3-70B2026.03 | 39 | 0 | |
| DataSentinelModel=Llama3.3-70B2026.03 | 33.79 | 19.95 | |
| SpotlightingModel=Gemini-2.52026.03 | 28.8 | 34.81 | |
| PI DetectorModel=Llama3.3-70B2026.03 | 28.46 | 5.9 | |
| DataSentinelModel=Gemini-2.52026.03 | 26.08 | 26.08 | |
| PromptGuardModel=Gemini-2.52026.03 | 24.94 | 19.84 | |
| MELONModel=Gemini-2.52026.03 | 21.43 | 0.23 | |
| PIGuardModel=Llama3.3-70B2026.03 | 15.87 | 0.11 | |
| PI DetectorModel=Gemini-2.52026.03 | 14.17 | 4.31 | |
| PIGuardModel=Gemini-2.52026.03 | 10.43 | 1.36 |