Jailbreak Attack on In-the-wild Jailbreak
60.3ASR (m)S-PIEL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| S-PIELTarget Model Type=Mistral 7B2025.03 | 60.3 | 62.7 | |
| S-PIELTarget Model Type=Gemma-2-9B2025.03 | 58.7 | 59.8 | |
| S-PIELTarget Model Type=Llama-2-7B2025.03 | 54.3 | 56.1 | |
| S-PIELTarget Model Type=Llama-3.1-8B2025.03 | 38.9 | 41 | |
| S-PIELTarget Model Type=Llama-3.1-8B (DeepSeek-R1-Distilled)2025.03 | 36.9 | 37.6 | |
| GCGTarget Model Type=Mistral 7B2025.03 | 19.7 | 20.3 | |
| GCGTarget Model Type=Llama-2-7B2025.03 | 18.9 | 20.1 | |
| GCGTarget Model Type=Gemma-2-9B2025.03 | 18.8 | 19.4 | |
| Vanilla PromptTarget Model Type=Mistral 7B2025.03 | 18.7 | 21.5 | |
| Vanilla PromptTarget Model Type=Gemma-2-9B2025.03 | 12.3 | 13.7 | |
| GCGTarget Model Type=Llama-3.1-8B2025.03 | 12.2 | 14.7 | |
| Vanilla PromptTarget Model Type=Llama-2-7B2025.03 | 12.1 | 14.4 | |
| GCGTarget Model Type=Llama-3.1-8B (DeepSeek-R1-Distilled)2025.03 | 8.9 | 9.7 | |
| Vanilla PromptTarget Model Type=Llama-3.1-8B2025.03 | 7.7 | 8.2 | |
| Vanilla PromptTarget Model Type=Llama-3.1-8B (DeepSeek-R1-Distilled)2025.03 | 6.5 | 6.9 |