Phare Framing Jailbreaks (English)
Safety · 2025-12-09
Phare jailbreak task using framing strategies to elicit unsafe behavior. This row is the English language split.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.5 | 98.2 |
| Opus 4.5 | 97.0 |
| GPT-5.2 | 96.9 |
| GPT-5 Mini | 96.9 |
| GPT-5 Nano | 95.7 |
| Kimi K2.6 | 95.7 |
| GPT-5 | 93.8 |
| Opus 4.6 | 93.4 |
| GPT-5.1 | 92.4 |
| Sonnet 4.5 | 91.0 |
| Sonnet 5 | 90.5 |
| Sonnet 4.6 | 90.5 |
| Opus 4.1 | 90.3 |
| Haiku 4.5 | 88.1 |
| Qwen3.7 Max Preview | 85.3 |
| Grok 4 | 83.3 |
| Grok 4.3 | 81.8 |
| Haiku 3.5 | 79.7 |
| Claude 3.7 Sonnet | 79.6 |
| Kimi K2.5 | 76.9 |
| Gemini 3 Pro Preview | 76.5 |
| Llama 3.1 405B Instruct | 76.2 |
| gpt-oss-120b | 75.4 |
| Qwen3.7-Plus | 74.7 |
| Gemma 4 31B IT | 74.6 |
| Gemini 3.1 Pro Preview | 73.0 |
| GLM-5.2 | 71.5 |
| Qwen3 Max (rolling alias) | 68.2 |
| Gemini 3.1 Flash-Lite | 68.0 |
| Gemini 3.5 Flash | 67.8 |
| Gemini 2.5 Flash-Lite | 64.7 |
| Qwen-Plus (2025-01-25) | 64.5 |
| GPT-4.1 Nano | 63.5 |
| Llama 3.1 8B Instruct | 62.7 |
| Grok 3 Mini | 60.2 |
| DeepSeek-V4-Flash | 60.1 |
| GPT-4.1 Mini | 60.1 |
| Qwen3 8B | 59.9 |
| GPT-4o Mini | 59.8 |
| GPT-4o | 59.6 |