Jailbreak Defense on ClearHarm (held-out)
47Attack Success Rate (ASR)MLPCT
Evaluation Results
| Method | Links | |
|---|---|---|
| MLPCTTrain threat=Sycophancy2026.06 | 47 | |
| Base modelNote=no training2026.06 | 49 | |
| BCTTrain threat=Jailbreak2026.06 | 51 | |
| BCTTrain threat=Frustration2026.06 | 87 |