List your company

Adversarial Attack on JailBreakBench (Qwen3-4B held-out subset)

30ASR (%)PGD-10
-1.26.91523.1May 14, 2026

Evaluation results

MethodLinks
PGD-10Gradient Type=Exact gradient2026.05300.60.18
PGD-5Gradient Type=Exact gradient2026.0561.20.07
FGSMGradient Type=Exact gradient2026.05460.24
PGD-2Gradient Type=Exact gradient2026.05430.12
RS-FGSMGradient Type=Exact gradient2026.0525.90.12
FGMGradient Type=Exact gradient2026.0505.60
FGSMGradient Type=GPR (predicted)2026.05059.40
FGMGradient Type=GPR (predicted)2026.05055.70
RS-FGSMGradient Type=GPR (predicted)2026.050590
PGD-2Gradient Type=GPR (predicted)2026.05031.20
PGD-5Gradient Type=GPR (predicted)2026.050130
PGD-10Gradient Type=GPR (predicted)2026.0506.60