Adversarial Robustness against Indirect Prompt Injection on AgentDojo Average across attacks
13.18UAPIGuard
Evaluation results
| Method | Links | ||
|---|---|---|---|
| PIGuardModel=Gemini-2.52026.03 | 13.18 | 0.85 | |
| PI DetectorModel=Gemini-2.52026.03 | 15.62 | 3.63 | |
| PIGuardModel=Llama3.3-70B2026.03 | 18.17 | 0.59 | |
| PI DetectorModel=Llama3.3-70B2026.03 | 24.55 | 3.89 | |
| DataSentinelModel=Gemini-2.52026.03 | 32.06 | 13.95 | |
| PromptGuardModel=Gemini-2.52026.03 | 34.58 | 9.63 | |
| MELONModel=Gemini-2.52026.03 | 34.7 | 0.26 | |
| DataSentinelModel=Llama3.3-70B2026.03 | 36.42 | 11.82 | |
| CaMeLModel=Llama3.3-70B2026.03 | 39.26 | 0 | |
| SpotlightingModel=Gemini-2.52026.03 | 41.58 | 18.26 | |
| PromptGuardModel=Llama3.3-70B2026.03 | 41.58 | 11.2 | |
| CaMeLModel=Gemini-2.52026.03 | 42.52 | 0 | |
| SandwichingModel=Llama3.3-70B2026.03 | 44.13 | 7.46 | |
| MELONModel=Llama3.3-70B2026.03 | 45.49 | 10.46 | |
| AttriGuardModel=Llama3.3-70B2026.03 | 48.02 | 0 | |
| SandwichingModel=Gemini-2.52026.03 | 48.5 | 15.19 | |
| PromptArmorModel=Gemini-2.52026.03 | 51.56 | 0.54 | |
| SpotlightingModel=Llama3.3-70B2026.03 | 52.18 | 14.06 | |
| AttriGuardModel=Gemini-2.52026.03 | 52.32 | 0 | |
| IPIGuardModel=Llama3.3-70B2026.03 | 52.58 | 2.21 | |
| PromptArmorModel=Llama3.3-70B2026.03 | 55.27 | 1.05 | |
| IPIGuardModel=Gemini-2.52026.03 | 66.61 | 3.43 |