Jailbreak Detection on MM-SafetyBench
99.18AUROCMahal-OOD
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Mahal-OODModel=LLaVA, Uses target model hidden states=true2025.12 | 99.18 | 99.32 | |
| Mahal-OOD*Model=FLAVA, Uses target model hidden states=false2025.12 | 97.01 | 97.62 | |
| HiddenDetectVLM=LLaVa2026.06 | 96.8 | — | |
| KNN-OODModel=LLaVA, Uses target model hidden states=true2025.12 | 95.23 | 95.91 | |
| JailDAM*Model=CLIP, Uses target model hidden states=false2025.12 | 91.26 | 98.04 | |
| MTKVLM=LLaVa2026.06 | 91.1 | — | |
| KNN-OOD*Model=FLAVA, Uses target model hidden states=false2025.12 | 85.47 | 90.65 | |
| GradSafeModel=LLaVA2025.12 | 85.14 | 87.52 | |
| ECSOVLM=LLaVa2026.06 | 84.8 | — | |
| HiddenDetectModel=LLaVA2025.12 | 82.69 | 93.53 | |
| MTKVLM=Qwen-VL2026.06 | 81.8 | — | |
| MirrorCheckVLM=Qwen-VL2026.06 | 78.1 | — | |
| HiddenDetectVLM=Qwen-VL2026.06 | 77.4 | — | |
| JailGuardVLM=Qwen-VL2026.06 | 75 | — | |
| LLaVaGuardModel=Qwen2025.12 | 74.27 | 87.29 | |
| MirrorCheckVLM=LLaVa2026.06 | 68.8 | — | |
| ECSOVLM=Qwen-VL2026.06 | 68.6 | — | |
| CIDERVLM=LLaVa2026.06 | 67.3 | — | |
| CIDERVLM=Qwen-VL2026.06 | 67.3 | — | |
| JailDAMVLM=LLaVa2026.06 | 61.2 | — | |
| VLGuardModel=LLaVA2025.12 | 61.06 | 80.2 | |
| JailGuardVLM=LLaVa2026.06 | 53.3 | — | |
| JailDAMVLM=Qwen-VL2026.06 | 51.9 | — |