System Prompt Extraction on attack prompts Out-of-domain
78ASR (TA)RogueMerge
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| RogueMerge2026.06 | 78 | 83.5 | 74 | 80 | 78.9 | |
| LoBAM*2026.06 | 33.5 | 45.5 | 42 | 36.5 | 39.4 | |
| MergeHijacking*2026.06 | 20 | 21 | 19.5 | 22 | 20.6 | |
| SFT2026.06 | 13.5 | 15 | 13.5 | 15.5 | 14.4 | |
| Clean2026.06 | 6 | 9.5 | 7 | 9.5 | 8 |