Jailbreak Attack on JailbreakBench (ASR, ASR@10)
81ASRFew-shot threat model
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Few-shot threat modelModel=GPT-OSS 120B, Steps=102026.05 | 81 | 100 | |
| Generation-phase threat modelModel=Qwen3 8B, Steps=102026.05 | 69.2 | 100 | |
| Generation-phase threat modelModel=GPT-OSS 120B, Steps=102026.05 | 67.6 | 98 | |
| Generation-phase threat modelModel=Qwen3 8B, Steps=52026.05 | 46.2 | 86 | |
| Generation-phase threat modelModel=GPT-OSS 120B, Steps=52026.05 | 39.6 | 88 | |
| Few-shot threat modelModel=GPT-OSS 120B, Steps=52026.05 | 26 | 86 | |
| Few-shot threat modelModel=Qwen3 8B, Steps=52026.05 | 17.8 | 56 | |
| Few-shot threat modelModel=Qwen3 8B, Steps=102026.05 | 0.2 | 2 | |
| Self-supervised threat modelModel=Qwen3 8B, Steps=52026.05 | 0 | 0 | |
| Self-supervised threat modelModel=Qwen3 8B, Steps=102026.05 | 0 | 0 | |
| Self-supervised threat modelModel=GPT-OSS 120B, Steps=52026.05 | 0 | 0 | |
| Self-supervised threat modelModel=GPT-OSS 120B, Steps=102026.05 | 0 | 0 |