Jailbreak Attack on HarmBench behaviors
100Attack Success Rate (ASR)I-GCG
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| I-GCGTarget Model=Qwen2.5-7B-Instruct2026.03 | 100 | 26 | |
| TAO-AttackTarget Model=Qwen2.5-7B-Instruct2026.03 | 100 | 11 | |
| TAO-AttackTarget Model=Llama-2-7B-Chat2026.03 | 90 | 357 | |
| I-GCGTarget Model=Llama-2-7B-Chat2026.03 | 45 | 792 |