List your company

Jailbreak Defense on JailbreakBench and AdvBench

0.1ASRCertified Semantic Smoothing
-3.26419.44342.1564.857Feb 2, 2026

Evaluation results

MethodLinks
Certified Semantic SmoothingAttack Algorithm=GCG, Retention rate (k)=0.32026.020.197.122.568.22.15
Certified Semantic SmoothingAttack Algorithm=TAP, Retention rate (k)=0.32026.020.197.422.968.22.15
Certified Semantic SmoothingAttack Algorithm=TAP, Retention rate (k)=0.52026.020.394.818.986.52.05
Certified Semantic SmoothingAttack Algorithm=GCG, Retention rate (k)=0.52026.020.594.218.486.52.05
Certified Semantic SmoothingAttack Algorithm=AutoDAN, Retention rate (k)=0.52026.020.693.518.186.52.05
Certified Semantic SmoothingAttack Algorithm=TAP, Retention rate (k)=0.72026.020.890.115.294.11.95
Certified Semantic SmoothingAttack Algorithm=GCG, Retention rate (k)=0.72026.021.289.414.694.11.95
Certified Semantic SmoothingAttack Algorithm=AutoDAN, Retention rate (k)=0.72026.021.588.714.394.11.95
SmoothLLMAttack Algorithm=GCG2026.023.562.14.274.32.1
SmoothLLMAttack Algorithm=TAP2026.024.161.84.174.32.1
SmoothLLMAttack Algorithm=AutoDAN2026.028.258.43.874.32.1
Erase-CheckAttack Algorithm=GCG2026.0212.4——88.54.2
Erase-CheckAttack Algorithm=TAP2026.0214.1——88.54.2
Self-DenoiseAttack Algorithm=GCG2026.0218.924.51.289.23.45
Self-DenoiseAttack Algorithm=TAP2026.0222.123.81.189.23.45
PPL FilterAttack Algorithm=TAP2026.0238.4——91.20.92
PPL FilterAttack Algorithm=AutoDAN2026.0241.2——91.20.92
PPL FilterAttack Algorithm=GCG2026.0245.3——91.20.92
VanillaAttack Algorithm=TAP (Black-box)2026.0268.90096.50.85
VanillaAttack Algorithm=AutoDAN2026.0272.10096.50.85
VanillaAttack Algorithm=GCG (White-box)2026.0284.20096.50.85