System Prompt Extraction on Attack Prompts In-domain
78Attack Success Rate (TA)RogueMerge
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| RogueMerge2026.06 | 78 | 80 | 76 | 78 | 78 | |
| LoBAM*2026.06 | 34 | 46 | 45 | 34 | 39.8 | |
| MergeHijacking*2026.06 | 16 | 17 | 15 | 17 | 16.3 | |
| SFT2026.06 | 10 | 13 | 7 | 14 | 11 | |
| Clean2026.06 | 9 | 7 | 8 | 7 | 7.8 |