Jailbreaking on HarmBench and StrongReject 200 prompts (held-out)
80Success Rate Fraction
PGD + Adaptive Probe-based Steering
Updated 13d ago
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| 80 | 97 | 98 | ||
| 71 | 85 | 97 | ||
| 67 | 94 | 97 | ||
| 66 | 88 | 96 | ||
| 66 | 81 | 92 | ||
2026.05 | 26 | 28 | 29 | |
2026.05 | 19 | 3 | 4 | |
2026.05 | 1 | 1 | 0 |