Prefill Attack Defense on ClearHarm prefill (held-out)
52PARBase model
Evaluation Results
| Method | Links | |
|---|---|---|
| Base modelNote=no training2026.06 | 52 | |
| BCTTrain threat=Frustration2026.06 | 40 | |
| MLPCTTrain threat=Sycophancy2026.06 | 34 | |
| BCTTrain threat=Jailbreak2026.06 | 34 |