Prompt Injection Attack Evaluation on JailbreakBench
100Attack Success Probability (ASP)Mistral
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MistralOpen-source=true2025.05 | 100 | 100 | |
| Neural-chatOpen-source=true2025.05 | 99.3 | 99.3 | |
| Stablelm2Open-source=true2025.05 | 97.3 | 95.3 | |
| OpenchatOpen-source=true2025.05 | 92 | 92 | |
| Starling-lmOpen-source=true2025.05 | 80.3 | 77.3 | |
| VicunaOpen-source=true2025.05 | 76 | 75.3 | |
| PhiOpen-source=true2025.05 | 64.3 | 59.3 | |
| Deepseek-r1Open-source=true2025.05 | 60 | 58 | |
| Gemma2Open-source=true2025.05 | 52.3 | 33.3 | |
| Phi3Open-source=true2025.05 | 35 | 33.3 | |
| Llama2Open-source=true2025.05 | 11.7 | 8 | |
| GemmaOpen-source=true2025.05 | 5.3 | 5.3 | |
| Llama3Open-source=true2025.05 | 4.7 | 4.7 | |
| Gemma-2bOpen-source=true2025.05 | 0.7 | 0.7 |