Jailbreaking on HarmBench Visual Text Replacement
3.6ASR (K=1)Claude Haiku 4.5
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Claude Haiku 4.52026.05 | 3.6 | 12.9 | |
| GPT-5.22026.05 | 9.4 | 14.4 | |
| Gemini 3 Flash2026.05 | 14.9 | 32.8 | |
| Qwen3-VL-235B2026.05 | 30.6 | 51.5 | |
| Gemini 3.1 Pro2026.05 | 32.6 | 48.6 | |
| Qwen3-VL-32B2026.05 | 36.8 | 58.1 |