Jailbreak Attack Success Rate Evaluation on jailbreak malicious pool (held-out)
53.52Attack Success Rate (ASR)Vanilla
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| VanillaModel=Gemma-2-9B, System=Vanilla2026.06 | 53.52 | 30.75 | 33.03 | |
| NeuroArmor-DetectModel=Gemma-2-9B, System=NeuroArmor-Detect2026.06 | 53.52 | 30.75 | 33.03 | |
| VanillaModel=Llama-3-8B, System=Vanilla2026.06 | 41.56 | 25.66 | 23.97 | |
| NeuroArmor-DetectModel=Llama-3-8B, System=NeuroArmor-Detect2026.06 | 41.56 | 25.66 | 23.97 | |
| Llama-Guard-3Model=Llama-3-8B, System=Llama-Guard-32026.06 | 26.61 | 16.97 | 10.56 | |
| Runtime Intervention BaselineModel=Llama-3-8B, System=Runtime Intervention Baseline2026.06 | 24.92 | 21.94 | 15.36 | |
| Direct Safety SteeringModel=Llama-3-8B, System=Direct Safety Steering2026.06 | 23.92 | 18.34 | 13.41 | |
| NeuroArmor-ReAnchorModel=Gemma-2-9B, System=NeuroArmor-ReAnchor2026.06 | 1.9 | 1.32 | 1.28 | |
| NeuroArmor-ReAnchorModel=Llama-3-8B, System=NeuroArmor-ReAnchor2026.06 | 1.57 | 0.87 | 0.99 |