Jailbreak Detection on SafeMTData
0.915AUCSAFEDREAM
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SAFEDREAMMethod category=Multi-turn stateful detectors (external), Target LLM=Qwen2.5-7B-Instruct2026.04 | 0.915 | 90.4 | 8 | 1.06 | |
| TARSMethod category=Alignment-based methods (modify LLM weights), Target LLM=Qwen2.5-7B-Instruct2026.04 | 0.847 | 83.2 | 4.2 | 0.31 | |
| X-BoundaryMethod category=Alignment-based methods (modify LLM weights), Target LLM=Qwen2.5-7B-Instruct2026.04 | 0.836 | 82 | 9.1 | 0.28 | |
| Circuit BreakerMethod category=Alignment-based methods (modify LLM weights), Target LLM=Qwen2.5-7B-Instruct2026.04 | 0.824 | 81.1 | 8.4 | 0.23 | |
| SelfDefendMethod category=external, stateless, Target LLM=Qwen2.5-7B-Instruct2026.04 | 0.718 | 68.5 | 23.8 | 0.35 | |
| GuardReasonerMethod category=external, stateless, Target LLM=Qwen2.5-7B-Instruct2026.04 | 0.691 | 53.4 | 13.9 | 0.39 | |
| BingoGuardMethod category=external, stateless, Target LLM=Qwen2.5-7B-Instruct2026.04 | 0.68 | 63.8 | 17.8 | 0.36 | |
| R2-GuardMethod category=external, stateless, Target LLM=Qwen2.5-7B-Instruct2026.04 | 0.675 | 59.1 | 13.7 | 0.41 | |
| WildGuardMethod category=external, stateless, Target LLM=Qwen2.5-7B-Instruct2026.04 | 0.672 | 57.7 | 19.4 | 0.68 |