Jailbreak Attack on harmful-query set
98.5Attack Success Rate (ASR)EvoSynth
Evaluation Results
| Method | Links | |
|---|---|---|
| EvoSynthDefense Method=Deliberative Alignment, Defense Category=White-box training, Target Model / Setup=OpenAI o12025.11 | 98.5 | |
| EvoSynthDefense Method=Circuit Breakers, Defense Category=White-box training, Target Model / Setup=Llama-3-8B-Circuit Breaker2025.11 | 95.6 | |
| EvoSynthDefense Method=SmoothLLM, Defense Category=Black-box defense layer, Target Model / Setup=All target models2025.11 | 95.2 | |
| EvoSynthDefense Method=SelfDefend, Defense Category=Black-box defense layer, Target Model / Setup=All target models2025.11 | 94.3 | |
| EvoSynthDefense Method=ProAct, Defense Category=Black-box defense layer, Target Model / Setup=All target models2025.11 | 92.5 | |
| EvoSynthDefense Method=EDDF, Defense Category=Black-box defense layer, Target Model / Setup=All target models2025.11 | 91.8 |