List your company

Prompt Injection Defense on Jailbreak APGD

1ASRGradient-guided Token Suppression
-2.8423.084974.92May 24, 2026

Evaluation results

MethodLinks
Gradient-guided Token SuppressionPerturbation limit (ε)=162026.051
CIDERPerturbation limit (ε)=162026.056
GUARDPerturbation limit (ε)=162026.056
GUARDPerturbation limit (ε)=322026.056
Gradient-guided Token SuppressionPerturbation limit (ε)=322026.056
ECSOPerturbation limit (ε)=322026.0513
ECSOPerturbation limit (ε)=162026.0516
CIDERPerturbation limit (ε)=322026.0527
ASTRAPerturbation limit (ε)=162026.0529
ASTRAPerturbation limit (ε)=322026.0532
W/O DEFENSEPerturbation limit (ε)=162026.0580
W/O DEFENSEPerturbation limit (ε)=322026.0597