Jailbreak Safety Evaluation on English dataset Single-Image
10StrongREJECT (Perturbed)Gemini 1.5 Pro
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Gemini 1.5 Pro2025.04 | 10 | 13 | 0 | |
| Gemini 1.5 Flash2025.04 | 8 | 39 | 19 | |
| Claude 3 Haiku2025.04 | 5 | 4 | 2 | |
| GPT-4o2025.04 | 1 | 10 | 9 | |
| Claude 3.5 Sonnet2025.04 | 0 | 0 | 0 | |
| GPT-4o Mini2025.04 | 0 | 1 | 10 |