Jailbreak Safety Evaluation on JailbreakBench
2Attack Success Rate (J)Direct
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DirectModel=Qwen2.5-7B-Instruct, Backdoor Trigger=None2026.04 | 2 | — | 13 | 6.4 | |
| τ_trigger ⊕ DirectModel=Qwen2.5-7B-Instruct, Backdoor Trigger=τ_trigger2026.04 | 18 | — | 10 | 8.8 | |
| PAIRModel=Qwen2.5-7B-Instruct, Backdoor Trigger=None2026.04 | 49 | — | 52 | 33.6 | |
| TAPModel=Qwen2.5-7B-Instruct, Backdoor Trigger=None2026.04 | 54 | — | 59.2 | 38.7 | |
| τ_trigger ⊕ DirectModel=Backdoor Model (ACB), Backdoor Trigger=τ_trigger2026.04 | 59 | — | 49 | 30.5 | |
| PAPModel=Qwen2.5-7B-Instruct, Backdoor Trigger=None2026.04 | 76 | — | 71 | 53.1 | |
| τ_trigger ⊕ TAPModel=Backdoor Model (ACB), Backdoor Trigger=τ_trigger2026.04 | 87.5 | — | 91 | 58.4 | |
| τ_trigger ⊕ PAIRModel=Backdoor Model (ACB), Backdoor Trigger=τ_trigger2026.04 | 89 | — | 90 | 58.1 | |
| τ_trigger ⊕ PAPModel=Backdoor Model (ACB), Backdoor Trigger=τ_trigger2026.04 | 90 | — | 82 | 56.8 | |
| QwQ-32BSetup=Original2026.03 | — | 23 | — | — | |
| R1-32BSetup=Original2026.03 | — | 27 | — | — | |
| SafeReActBackbone=QwQ-32B2026.03 | — | 2 | — | — | |
| SafeReActBackbone=R1-32B2026.03 | — | 3 | — | — |