List your company

Prompt Injection Defense on Jailbreak MI-FGSM

4Attack Success RateGUARD
0.3225.165074.84May 24, 2026

Evaluation results

MethodLinks
GUARDPerturbation limit (ε)=162026.054
Gradient-guided Token SuppressionPerturbation limit (ε)=162026.054
Gradient-guided Token SuppressionPerturbation limit (ε)=322026.055
CIDERPerturbation limit (ε)=162026.056
GUARDPerturbation limit (ε)=322026.057
CIDERPerturbation limit (ε)=322026.0516
ECSOPerturbation limit (ε)=162026.0517
ECSOPerturbation limit (ε)=322026.0518
ASTRAPerturbation limit (ε)=322026.0533
ASTRAPerturbation limit (ε)=162026.0542
W/O DEFENSEPerturbation limit (ε)=322026.0595
W/O DEFENSEPerturbation limit (ε)=162026.0596