Jailbreak on AdvBench 50 most harmful requests
95.83Attack Success Rate (ASR)MIDAS
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MIDASTarget Model=QVQ-Max2026.02 | 95.83 | 4.19 | |
| MIDASTarget Model=Gemini-2.5-FT2026.02 | 90 | 4.57 | |
| MIDASTarget Model=GPT-4o2026.02 | 80 | 3.12 | |
| MIDASTarget Model=GPT-5-Chat2026.02 | 64 | 3.02 | |
| FigStepTarget Model=QVQ-Max2026.02 | 30.61 | 1.04 | |
| HIMRDTarget Model=Gemini-2.5-FT2026.02 | 18.3 | 0.71 | |
| HIMRDTarget Model=GPT-4o2026.02 | 12 | 0.54 | |
| HIMRDTarget Model=QVQ-Max2026.02 | 10.2 | 0.67 | |
| FigStepTarget Model=Gemini-2.5-FT2026.02 | 0 | 0.6 | |
| FigStepTarget Model=GPT-4o2026.02 | 0 | 0.35 | |
| FigStepTarget Model=GPT-5-Chat2026.02 | 0 | 0 | |
| HIMRDTarget Model=GPT-5-Chat2026.02 | 0 | 0.08 |