Indirect Prompt Injection Defense on Vision-Language Agentic IPI Benchmark (test)
72BUNo Defense
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| No DefenseFoundation Model=MINICPM2026.02 | 72 | 59.7 | 29 | |
| Repeat PromptFoundation Model=MINICPM2026.02 | 72 | 59.6 | 22 | |
| LLM-Detector (Gemini)Foundation Model=MINICPM2026.02 | 72 | 0 | 30.3 | |
| ICONFoundation Model=MINICPM2026.02 | 72 | 2.4 | 48.8 | |
| No DefenseFoundation Model=QWEN3VL2026.02 | 70.5 | 69.9 | 21.3 | |
| Repeat PromptFoundation Model=QWEN3VL2026.02 | 70.5 | 47.8 | 46.5 | |
| LLM-Detector (Gemini)Foundation Model=QWEN3VL2026.02 | 70.5 | 0 | 21.9 | |
| ICONFoundation Model=QWEN3VL2026.02 | 70.5 | 3.8 | 44.9 | |
| No DefenseFoundation Model=INTERNVL2026.02 | 64.5 | 47.4 | 33.9 | |
| Repeat PromptFoundation Model=INTERNVL2026.02 | 64.5 | 41.6 | 26.5 | |
| LLM-Detector (Gemini)Foundation Model=INTERNVL2026.02 | 64.5 | 0 | 34.1 | |
| ICONFoundation Model=INTERNVL2026.02 | 64.5 | 2.6 | 47.8 |