Jailbreaking on HarmBench FigStep
3ASR (K=1)GPT-5.2
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GPT-5.22026.05 | 3 | 3.8 | |
| Gemini 3.1 Pro2026.05 | 4 | 10.1 | |
| Gemini 3 Flash2026.05 | 10.1 | 10.1 | |
| Qwen3-VL-32B2026.05 | 11.3 | 11.3 | |
| Claude Haiku 4.52026.05 | 33 | 45.9 | |
| Qwen3-VL-235B2026.05 | 37 | 49.1 |