Indirect Prompt Injection Defense on TrojanTools
47.1ASRNo Defense
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| No DefenseFoundation Model=LLAMA-3.1-8B, Base Utility=34.52026.02 | 47.1 | 14.3 | |
| No DefenseFoundation Model=MISTRAL-8B, Base Utility=20.12026.02 | 42.9 | 8.5 | |
| No DefenseFoundation Model=QWEN-3-8B, Base Utility=69.82026.02 | 42.7 | 35.3 | |
| DelimitingFoundation Model=QWEN-3-8B, Base Utility=69.82026.02 | 31.8 | 43.7 | |
| DelimitingFoundation Model=LLAMA-3.1-8B, Base Utility=34.52026.02 | 30.5 | 24.6 | |
| Repeat PromptFoundation Model=LLAMA-3.1-8B, Base Utility=34.52026.02 | 28.9 | 25.5 | |
| DelimitingFoundation Model=MISTRAL-8B, Base Utility=20.12026.02 | 27.9 | 10.2 | |
| Repeat PromptFoundation Model=QWEN-3-8B, Base Utility=69.82026.02 | 25.4 | 45.6 | |
| Repeat PromptFoundation Model=MISTRAL-8B, Base Utility=20.12026.02 | 23.2 | 12.3 | |
| MELONFoundation Model=QWEN-3-8B, Base Utility=69.82026.02 | 5.8 | 36.8 | |
| MELONFoundation Model=LLAMA-3.1-8B, Base Utility=34.52026.02 | 3.5 | 14.8 | |
| MELONFoundation Model=MISTRAL-8B, Base Utility=20.12026.02 | 2.6 | 9.3 | |
| ICONFoundation Model=QWEN-3-8B, Base Utility=69.82026.02 | 1.2 | 50.5 | |
| ICONFoundation Model=MISTRAL-8B, Base Utility=20.12026.02 | 0.9 | 13.9 | |
| ICONFoundation Model=LLAMA-3.1-8B, Base Utility=34.52026.02 | 0.8 | 27.1 | |
| LLM-DetectorFoundation Model=QWEN-3-8B, Base Utility=69.82026.02 | 0.6 | 38.6 | |
| LLM-DetectorFoundation Model=LLAMA-3.1-8B, Base Utility=34.52026.02 | 0 | 15.1 | |
| LLM-DetectorFoundation Model=MISTRAL-8B, Base Utility=20.12026.02 | 0 | 9.7 |