Jailbreak Defense on Jailbreak Attack Suite (GCG, AutoDAN, DeepInc., Prefill, Int.Laund.)
55GCG Defense RateDual-probe gated
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Dual-probe gatedModel=Mistral-7B, Paradigm=Dual-probe gated (ours)2026.06 | 55 | 93 | 8 | 8 | 15 | 62.2 | |
| No defenseModel=Mistral-7B, Paradigm=No defense2026.06 | 30 | 95 | 10 | 68 | 20 | 52.8 | |
| Probe-gated (prompt)Model=Mistral-7B, Paradigm=Probe-gated (prompt)2026.06 | 20 | 95 | 3 | 73 | 20 | 55.6 | |
| CASTModel=Mistral-7B, Paradigm=CAST2026.06 | 13 | 98 | 3 | 63 | 15 | 59.4 | |
| AlphaSteerModel=Qwen-2.5-7B, Paradigm=AlphaSteer2026.06 | 8 | 5 | 18 | 60 | 0 | 80 | |
| AlphaSteerModel=Gemma-2-9B, Paradigm=AlphaSteer2026.06 | 5 | 75 | 0 | 33 | 10 | 73.9 | |
| StaticModel=Qwen-2.5-7B, Paradigm=Static2026.06 | 5 | 5 | 15 | 60 | 10 | 80 | |
| Probe-gated (prompt)Model=Llama-3.1-8B, Paradigm=Probe-gated (prompt)2026.06 | 5 | 3 | 3 | 55 | 15 | 83.9 | |
| StaticModel=Mistral-7B, Paradigm=Static2026.06 | 3 | 73 | 3 | 53 | 20 | 68.9 | |
| No defenseModel=Qwen-2.5-7B, Paradigm=No defense2026.06 | 3 | 8 | 15 | 65 | 0 | 80 | |
| CASTModel=Qwen-2.5-7B, Paradigm=CAST2026.06 | 3 | 5 | 15 | 65 | 0 | 80.6 | |
| Probe-gated (prompt)Model=Qwen-2.5-7B, Paradigm=Probe-gated (prompt)2026.06 | 3 | 5 | 15 | 63 | 10 | 80 | |
| No defenseModel=Llama-3.1-8B, Paradigm=No defense2026.06 | 3 | 3 | 5 | 53 | 30 | 82.8 | |
| CASTModel=Llama-3.1-8B, Paradigm=CAST2026.06 | 3 | 5 | 5 | 58 | 40 | 80 | |
| AlphaSteerModel=Mistral-7B, Paradigm=AlphaSteer2026.06 | 0 | 0 | 0 | 50 | 0 | 88.9 | |
| Unified (AlphaSteer + halt)Model=Mistral-7B, Paradigm=Unified (AlphaSteer + halt, ours)2026.06 | 0 | 0 | 0 | 8 | 0 | 98.3 | |
| No defenseModel=Gemma-2-9B, Paradigm=No defense2026.06 | 0 | 65 | 3 | 28 | 15 | 77.2 | |
| StaticModel=Gemma-2-9B, Paradigm=Static2026.06 | 0 | 65 | 5 | 28 | 10 | 77.2 | |
| CASTModel=Gemma-2-9B, Paradigm=CAST2026.06 | 0 | 65 | 5 | 28 | 10 | 77.2 | |
| Probe-gated (prompt)Model=Gemma-2-9B, Paradigm=Probe-gated (prompt)2026.06 | 0 | 65 | 3 | 28 | 10 | 77.8 | |
| StaticModel=Llama-3.1-8B, Paradigm=Static2026.06 | 0 | 0 | 0 | 75 | 15 | 81.7 | |
| AlphaSteerModel=Llama-3.1-8B, Paradigm=AlphaSteer2026.06 | 0 | 0 | 3 | 50 | 0 | 88.3 | |
| Dual-probe gatedModel=Llama-3.1-8B, Paradigm=Dual-probe gated (ours)2026.06 | 0 | 3 | 5 | 0 | 15 | 96.7 | |
| Unified (AlphaSteer + halt)Model=Llama-3.1-8B, Paradigm=Unified (AlphaSteer + halt, ours)2026.06 | 0 | 0 | 3 | 0 | 0 | 99.4 |