Red-teaming on 50 harmful goals (Manual evaluation)
100Hard ASRPAIR
Evaluation Results
| Method | Links | |
|---|---|---|
| PAIRTarget Model=Gemini 2.5 Flash2026.03 | 100 | |
| ITMS-VDTarget Model=Qwen2.5-Omni2026.03 | 100 | |
| PAIRTarget Model=GPT-4o2026.03 | 98 | |
| ITMS-CrescendoTarget Model=Gemini 2.5 Flash2026.03 | 98 | |
| CrescendoTarget Model=Gemini 3 Flash2026.03 | 98 | |
| PAIRTarget Model=Qwen2.5-Omni2026.03 | 98 | |
| CrescendoTarget Model=Qwen3-Omni2026.03 | 98 | |
| CrescendoTarget Model=GPT-4o2026.03 | 96 | |
| PAIRTarget Model=Gemini 3 Flash2026.03 | 96 | |
| CrescendoTarget Model=Qwen2.5-Omni2026.03 | 96 | |
| PAIRTarget Model=Qwen3-Omni2026.03 | 96 | |
| CrescendoTarget Model=Gemini 2.5 Flash2026.03 | 94 | |
| ITMS-CrescendoTarget Model=Gemini 3 Flash2026.03 | 94 | |
| ITMS-CrescendoTarget Model=Qwen3-Omni2026.03 | 94 | |
| ITMS-CrescendoTarget Model=Claude Sonnet 42026.03 | 92 | |
| ITMS-CrescendoTarget Model=GPT-4o2026.03 | 92 | |
| CrescendoTarget Model=Claude Sonnet 42026.03 | 90 | |
| ITMS-CrescendoTarget Model=Qwen2.5-Omni2026.03 | 88 | |
| Violent DurianTarget Model=Qwen2.5-Omni2026.03 | 86 | |
| ITMS-VDTarget Model=Gemini 2.5 Flash2026.03 | 62 | |
| PAIRTarget Model=Claude Sonnet 42026.03 | 60 | |
| Violent DurianTarget Model=Gemini 2.5 Flash2026.03 | 56 | |
| Violent DurianTarget Model=GPT-4o2026.03 | 42 | |
| ITMS-VDTarget Model=GPT-4o2026.03 | 40 | |
| ITMS-VDTarget Model=Gemini 3 Flash2026.03 | 34 | |
| Violent DurianTarget Model=Qwen3-Omni2026.03 | 30 | |
| Violent DurianTarget Model=Gemini 3 Flash2026.03 | 26 | |
| ITMS-VDTarget Model=Qwen3-Omni2026.03 | 22 | |
| ITMS-VDTarget Model=Claude Sonnet 42026.03 | 6 | |
| Violent DurianTarget Model=Claude Sonnet 42026.03 | 2 |