Jailbreak Attack Success Rate by Attack Method on AdvBench
99.5Safe2Harm ASR (Answer)Gemini 2.5
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemini 2.5Thinking Mode=false2025.12 | 99.5 | — | 44.5 | — | 0.82 | — | 5 | — | 35 | — | 34 | — | |
| DeepSeekThinking Mode=false2025.12 | 98 | — | 86.5 | — | 4.24 | — | 37.5 | — | 77 | — | 89.5 | — | |
| Qwen3 8BThinking Mode=true2025.12 | 92 | 100 | 82 | 96.5 | 3.28 | 92.62 | 45.19 | 98.08 | 90.34 | 92.23 | 90 | 91 | |
| Qwen3 8BThinking Mode=false2025.12 | 89 | — | 65.33 | — | 0 | — | 23.85 | — | 83.9 | — | 70.5 | — | |
| Qwen3 4BThinking Mode=true2025.12 | 87.5 | 100 | 88 | 95 | 4.1 | 97.54 | 72.88 | 78.65 | 91.86 | 92.05 | 90 | 89 | |
| Llama-3-8B-InstructThinking Mode=false2025.12 | 87 | — | 35.5 | — | 20.49 | — | 18.65 | — | 31.44 | — | 15 | — | |
| Qwen3 1.7BThinking Mode=false2025.12 | 85.93 | — | 88.5 | — | 6.56 | — | 83.08 | — | 87.31 | — | 70.5 | — | |
| Qwen3 1.7BThinking Mode=true2025.12 | 81 | 64.5 | 91.5 | 90.5 | 28.69 | 95.08 | 91.54 | 98.46 | 92.38 | 91.29 | 75 | 86 | |
| Qwen3 4BThinking Mode=false2025.12 | 79 | — | 69 | — | 0.82 | — | 51.54 | — | 85.41 | — | 87.44 | — | |
| GPT-5Thinking Mode=false2025.12 | 46 | — | 57.5 | — | 5.74 | — | 52 | — | 29 | — | 5 | — |