List your company

Jailbreak Attack on Qwen 27B 3.6

89ASRVanilla AutoDAN
-3.5620.4744.568.53May 27, 2026

Evaluation results

MethodLinks
Vanilla AutoDANAttack Type=Gradient-based search2026.058917.44
Block 10 LR ProbeMethod=Mechanistic AutoDAN, Probe Type=Logistic Regression, Block=102026.05824.79
Block 29 MLP ProbeMethod=Mechanistic AutoDAN, Probe Type=Multi-Layer Perceptron, Block=292026.05827.05
Block 10 MLP ProbeMethod=Mechanistic AutoDAN, Probe Type=Multi-Layer Perceptron, Block=102026.05797.05
Block 32 LR ProbeMethod=Mechanistic AutoDAN, Probe Type=Logistic Regression, Block=322026.05738.18
Block 1 LR ProbeMethod=Mechanistic AutoDAN, Probe Type=Logistic Regression, Block=12026.05321.77
Block 1 MLP ProbeMethod=Mechanistic AutoDAN, Probe Type=Multi-Layer Perceptron, Block=12026.05281.47
Template OnlyAttack Type=Baseline2026.050—