List your company

Safety Defense Evaluation on Jailbreak Attacks (GCG, AutoDAN, PAIR, DeepInc., SAP30, SIJ)

1GCG Attack ScorePPL
0.9881.0691.151.231Oct 23, 2025

Evaluation results

MethodLinks
PPLBackbone=Vicuna-13B2025.1014.323.13.13.922.763.03
Self-ExaminationBackbone=Vicuna-13B2025.1013.31.962.681.71.862.08
SAIDBackbone=Vicuna-13B2025.10111.1611.0311.03
No DefenseBackbone=Vicuna-13B2025.101.144.33.23.143.952.583.05
ParaphraseBackbone=Vicuna-13B2025.101.142.961.963.462.71.562.3
ICDBackbone=Vicuna-13B2025.101.144.042.483.242.553.162.77
SafeDecodingBackbone=Vicuna-13B2025.101.241.321.81.21.743.621.82
RetokenizationBackbone=Vicuna-13B2025.101.33.63.822.384.463.163.12