Indirect Prompt Injection Defense on Audio Modality (test)
7.5UIAinjectSystem Prompt
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| System Prompt2025.12 | 7.5 | 63.8 | 27.9 | 34.4 | 5 | |
| Ignore Prompt2025.12 | 24.3 | 65.7 | 28 | 34.7 | 2 | |
| Noise2025.12 | 42.8 | 41 | 0 | 0 | 2 | |
| No Defense2025.12 | 45.6 | 65.7 | 12.6 | 16.6 | 0 | |
| ARGUSsearch=disabled2025.12 | 54.4 | 65.7 | 0 | 0 | 4 | |
| AT2025.12 | 55.8 | 60.9 | 1.4 | 1.6 | 0 | |
| ARGUSadaptive intervention=disabled2025.12 | 57.2 | 65.7 | 0 | 0 | 3 | |
| ARGUSvariant=full2025.12 | 58 | 65.7 | 0 | 0 | 4 | |
| ARGUSpost-filtering=disabled2025.12 | 58.2 | 65.7 | 1 | 1 | 4 |