Adversarial Attack on AdvBench 50 harmful behaviors
100ASRGPTFuzzer
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GPTFuzzerTarget Model=Vicuna-7B2025.06 | 100 | — | |
| Algorithm 1 (Greedy Random Search)Target Model=Mistral 7B2026.05 | 90 | 18,271.92 | |
| VERATarget Model=Vicuna-7B2025.06 | 86 | — | |
| AutoDANTarget Model=Vicuna-7B2025.06 | 82 | — | |
| GCGTarget Model=Vicuna-7B2025.06 | 72 | — | |
| Algorithm 1 (Greedy Random Search)Target Model=Qwen 2.5 7B2026.05 | 64 | 38,229.16 | |
| PAIRTarget Model=Vicuna-7B2025.06 | 64 | — | |
| Algorithm 1 (Greedy Random Search)Target Model=LLaMA 2 Chat 7B2026.05 | 62 | 33,536.2 | |
| Algorithm 1 (Greedy Random Search)Target Model=Gemma 7B2026.05 | 52 | 49,182.39 | |
| VERATarget Model=LLaMA2-7B-Chat2025.06 | 16 | — | |
| AutoDANTarget Model=LLaMA2-7B-Chat2025.06 | 12 | — | |
| GPTFuzzerTarget Model=LLaMA2-7B-Chat2025.06 | 12 | — | |
| GCGTarget Model=LLaMA2-7B-Chat2025.06 | 10 | — | |
| PAIRTarget Model=LLaMA2-7B-Chat2025.06 | 8 | — |