List your company

Jailbreak Robustness on HarmBench 1.0 (test)

3.13GCG Attack Success RateCircuit-Breaker (CB)
0.721616.978333.23549.4917May 23, 2026

Evaluation results

MethodLinks
Circuit-Breaker (CB)Base Model=LLaMA-3-8B, Steering Method=CB2026.053.131.213.085.896.563.263.85
bi-level adversarial training frameworkBase Model=LLaMA-3-8B, Steering Method=Ours2026.054.311.372.16.076.313.94.01
bi-level adversarial training frameworkBase Model=Qwen-2.5-7B, Steering Method=Ours2026.054.821.593.955.866.054.54.46
Alpha-steerBase Model=Qwen-2.5-7B, Steering Method=AlphaSteer2026.055.211.734.845.215.664.224.48
Circuit-Breaker (CB)Base Model=Qwen-2.5-7B, Steering Method=CB2026.055.471.54.48.869.494.465.7
Latent Adversarial Training (LAT)Base Model=Qwen-2.5-7B, Steering Method=LAT2026.055.91.925.557.78.045.785.82
Alpha-steerBase Model=LLaMA-3-8B, Steering Method=AlphaSteer2026.056.621.453.678.237.784.535.38
ROSIBase Model=Qwen-2.5-7B, Steering Method=ROSI2026.056.792.126.8711.4212.756.877.8
bi-level adversarial training frameworkBase Model=Mistral-v2-7B, Steering Method=Ours2026.057.11.353.216.316.696.925.26
Circuit-Breaker (CB)Base Model=Mistral-v2-7B, Steering Method=CB2026.057.691.233.566.547.486.025.42
Latent Adversarial Training (LAT)Base Model=LLaMA-3-8B, Steering Method=LAT2026.057.91.122.547.187.823.755.05
Alpha-steerBase Model=Mistral-v2-7B, Steering Method=AlphaSteer2026.058.221.843.47.777.26.755.86
ROSIBase Model=LLaMA-3-8B, Steering Method=ROSI2026.059.341.683.799.48.643.46.04
Latent Adversarial Training (LAT)Base Model=Mistral-v2-7B, Steering Method=LAT2026.059.431.45.588.438.277.16.7
ROSIBase Model=Mistral-v2-7B, Steering Method=ROSI2026.0510.562.257.7610.610.329.348.47
LLaMA-3-8BBase Model=LLaMA-3-8B, Configuration=Base2026.0529.793.0411.4415.9617.212.1614.93
Qwen-2.5-7BBase Model=Qwen-2.5-7B, Configuration=Base2026.0549.9438.8125.6832.6234.3118.2733.27
Mistral-v2-7BBase Model=Mistral-v2-7B, Configuration=Base2026.0563.3474.6751.3462.8665.4940.5959.72