Adversarial Attack Transfer on 388 harmful behaviors
81.19Attack Success Rate (ASR)I-GCG
Evaluation Results
| Method | Links | |
|---|---|---|
| I-GCGTarget Model=Vicuna 7B 1.5v2026.06 | 81.19 | |
| Universal SlotGCGBase Attack=I-GCG, Target Model=Vicuna 7B 1.5v2026.06 | 79.9 | |
| Universal SlotGCGBase Attack=AttnGCG, Target Model=Vicuna 7B 1.5v2026.06 | 76.8 | |
| GCGTarget Model=Vicuna 7B 1.5v2026.06 | 70.1 | |
| AttnGCGTarget Model=Vicuna 7B 1.5v2026.06 | 64.43 | |
| GCG-HijTarget Model=Vicuna 7B 1.5v2026.06 | 63.66 | |
| Universal SlotGCGBase Attack=GCG, Target Model=Vicuna 7B 1.5v2026.06 | 63.4 | |
| Universal SlotGCGBase Attack=GCG-Hij, Target Model=Vicuna 7B 1.5v2026.06 | 61.34 | |
| Universal SlotGCGBase Attack=GCG, Target Model=GPT-3.5-turbo2026.06 | 50.77 | |
| Universal SlotGCGBase Attack=AttnGCG, Target Model=GPT-3.5-turbo2026.06 | 43.04 | |
| Universal SlotGCGBase Attack=GCG-Hij, Target Model=GPT-3.5-turbo2026.06 | 41.49 | |
| Universal SlotGCGBase Attack=I-GCG, Target Model=GPT-3.5-turbo2026.06 | 40.46 | |
| GCG-HijTarget Model=GPT-3.5-turbo2026.06 | 37.89 | |
| AttnGCGTarget Model=GPT-3.5-turbo2026.06 | 25.52 | |
| Universal SlotGCGBase Attack=I-GCG, Target Model=Average2026.06 | 24.83 | |
| Universal SlotGCGBase Attack=GCG, Target Model=Average2026.06 | 24.73 | |
| Universal SlotGCGBase Attack=AttnGCG, Target Model=Average2026.06 | 24.68 | |
| Universal SlotGCGBase Attack=GCG-Hij, Target Model=Average2026.06 | 22.38 | |
| GCG-HijTarget Model=Average2026.06 | 20.62 | |
| I-GCGTarget Model=Average2026.06 | 18.92 | |
| AttnGCGTarget Model=Average2026.06 | 18.2 | |
| GCGTarget Model=Average2026.06 | 14.9 | |
| I-GCGTarget Model=GPT-3.5-turbo2026.06 | 12.89 | |
| Universal SlotGCGBase Attack=GCG-Hij, Target Model=Gemini 2.5 Pro2026.06 | 6.7 | |
| Universal SlotGCGBase Attack=GCG-Hij, Target Model=Gemini 2.0 Flash2026.06 | 4.12 | |
| Universal SlotGCGBase Attack=GCG, Target Model=Gemini 2.5 Pro2026.06 | 3.61 | |
| GCGTarget Model=GPT-3.5-turbo2026.06 | 3.09 | |
| Universal SlotGCGBase Attack=GCG, Target Model=Gemini 2.0 Flash2026.06 | 2.06 | |
| Universal SlotGCGBase Attack=GCG, Target Model=GPT-4o2026.06 | 1.8 | |
| Universal SlotGCGBase Attack=AttnGCG, Target Model=GPT-4o2026.06 | 1.55 | |
| GCG-HijTarget Model=Gemini 2.5 Pro2026.06 | 1.55 | |
| GCGTarget Model=Gemini 2.0 Flash2026.06 | 1.29 | |
| Universal SlotGCGBase Attack=I-GCG, Target Model=GPT-4o2026.06 | 0.77 | |
| Universal SlotGCGBase Attack=I-GCG, Target Model=Gemini 2.0 Flash2026.06 | 0.77 | |
| AttnGCGTarget Model=GPT-4o2026.06 | 0.52 | |
| I-GCGTarget Model=Gemini 2.5 Pro2026.06 | 0.52 | |
| Universal SlotGCGBase Attack=GCG-Hij, Target Model=GPT-4o2026.06 | 0.26 | |
| AttnGCGTarget Model=Gemini 2.0 Flash2026.06 | 0.26 | |
| AttnGCGTarget Model=Gemini 2.5 Pro2026.06 | 0.26 | |
| Universal SlotGCGBase Attack=I-GCG, Target Model=Gemini 2.5 Pro2026.06 | 0.26 | |
| GCGTarget Model=GPT-4o2026.06 | 0 | |
| I-GCGTarget Model=GPT-4o2026.06 | 0 | |
| GCG-HijTarget Model=GPT-4o2026.06 | 0 | |
| Universal SlotGCGBase Attack=AttnGCG, Target Model=Gemini 2.0 Flash2026.06 | 0 | |
| I-GCGTarget Model=Gemini 2.0 Flash2026.06 | 0 | |
| GCG-HijTarget Model=Gemini 2.0 Flash2026.06 | 0 | |
| GCGTarget Model=Gemini 2.5 Pro2026.06 | 0 | |
| Universal SlotGCGBase Attack=AttnGCG, Target Model=Gemini 2.5 Pro2026.06 | 0 |