List your company

Jailbreak Detection on Jailbreak Detection Evaluation Set (combined)

0.895AccuracyKCD
0.586120.666310.74650.82669Dec 12, 2025

Evaluation results

MethodLinks
KCDLayer=21, Target Model=InternVL3-8B2025.120.8950.9750.1550.90.9260.924
MCDLayer=21, Target Model=InternVL3-8B2025.120.8910.9520.170.8920.9620.96
MCDLayer=20, Target Model=InternVL3-8B2025.120.8870.9750.2010.8910.950.94
KCDLayer=22, Target Model=InternVL3-8B2025.120.8840.9760.2080.9220.9250.922
MCDLayer=22, Target Model=InternVL3-8B2025.120.880.960.20.9160.960.959
KCDLayer=20, Target Model=InternVL3-8B2025.120.8770.970.2160.8980.930.929
JailDAM (RCS)Layer=CLIP, Target Model=Target Model-Agnostic2025.120.8450.9340.2440.8580.9150.9
JailDAM (Original)Layer=CLIP, Target Model=Target Model-Agnostic2025.120.7170.7060.2710.7140.7890.826
JailGuardLayer=Output, Target Model=InternVL3-8B2025.120.6920.60.2160.6610.7620.775
GradSafeLayer=Critical Param., Target Model=InternVL3-8B2025.120.6910.9290.5470.7490.8020.795
HiddenDetectLayer=18–24, Target Model=InternVL3-8B2025.120.6250.6870.5710.8120.7590.784
K-MeansLayer=FLAVA, Target Model=Target Model-Agnostic2025.120.6230.9740.7290.7210.5980.62
LogisticLayer=FLAVA, Target Model=Target Model-Agnostic2025.120.5980.3760.180.4830.6170.668