Prompt Injection Attack Robustness on HarmBench 2024 (test)
65.6Attack Success Probability (ASP)Role-playing CoT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Role-playing CoTAttack Method=Role-playing CoT, Model Group=Open-source LLMs (Mean)2025.05 | 65.6 | 64.6 | |
| HypnotismAttack Method=Hypnotism, Model Group=Open-source LLMs (Mean)2025.05 | 62.1 | 60.3 | |
| Ignore PrefixAttack Method=Ignore Prefix, Model Group=Open-source LLMs (Mean)2025.05 | 59.4 | 56.9 |