Jailbreak Attack on LegalQA
6Attack Success Rate (ASR)H-COT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| H-COTVictim LRM=OpenAI o4-mini2026.04 | 6 | 2 | |
| H-COTVictim LRM=DeepSeek R12026.04 | 7 | 2.1 | |
| H-COTVictim LRM=Qwen2.5-Max2026.04 | 7 | 2 | |
| Cognitive OverloadVictim LRM=Qwen2.5-Max2026.04 | 8 | 2 | |
| Cognitive OverloadVictim LRM=OpenAI o4-mini2026.04 | 8 | 2 | |
| Cognitive OverloadVictim LRM=DeepSeek R12026.04 | 10 | 2.17 | |
| PRJAVictim LRM=OpenAI o4-mini2026.04 | 65 | 2.34 | |
| PRJAVictim LRM=DeepSeek R12026.04 | 82 | 2.67 | |
| PRJAVictim LRM=Qwen2.5-Max2026.04 | 84 | 2.51 |