Directional Jailbreak Detection on WildJailbreak (Safe) × HarmBench (Harmful)
1AUROC (mono)Qwen3-8B
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen3-8BProbe Layer=L25, Relative Depth=~69%, Total Layers=36, Direction=downward2026.06 | 1 | 0.83 | 0.843 | |
| Llama-3.1-8BProbe Layer=L22, Relative Depth=~69%, Total Layers=32, Direction=downward2026.06 | 0.996 | 0.835 | 0.842 | |
| Gemma-7bProbe Layer=L19, Relative Depth=~68%, Total Layers=28, Direction=upward2026.06 | 0.78 | 0.884 | 0.889 |