Jailbreak Detection on GCG, Dolly 15K, and OR-Bench PMPs (test)
99True Positive Rate (TPR)SelfDefend
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SelfDefendSetting=Vanilla2026.06 | 99 | 4 | 55 | 85 | |
| SelfDefendSetting=+PMP Augment2026.06 | 99 | 3 | 26 | 86 | |
| SmoothLLMSetting=Vanilla2026.06 | 95 | 6 | 4 | 95 | |
| MTK2026.06 | 95 | 5 | 2 | 95 | |
| GradCuffSetting=Vanilla2026.06 | 90 | 5 | 12 | 95 | |
| GradCuffSetting=+PMP Augment2026.06 | 85 | 4 | 8 | 90 | |
| HiddenDetectSetting=Vanilla2026.06 | 83 | 5 | 6 | 88 | |
| SmoothLLMSetting=+PMP Augment2026.06 | 81 | 6 | 2 | 87 | |
| HiddenDetectSetting=+PMP Augment2026.06 | 80 | 3 | 4 | 87 | |
| SaPSetting=Vanilla2026.06 | 73 | 4 | 3 | 83 | |
| GradSafeSetting=Vanilla2026.06 | 71 | 4 | 9 | 78 | |
| GradSafeSetting=+PMP Augment2026.06 | 67 | 5 | 8 | 75 | |
| SaPSetting=+PMP Augment2026.06 | 41 | 4 | 1 | 58 | |
| HSFSetting=Vanilla2026.06 | 38 | 5 | 32 | 53 | |
| HSFSetting=+PMP Augment2026.06 | 23 | 4 | 27 | 38 |