Phare Framing Jailbreaks
Safety · 2025-12-09
Phare jailbreak task using framing strategies to elicit unsafe behavior.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.5 | 99.3 |
| Opus 4.5 | 98.9 |
| GPT-5 Nano | 98.5 |
| GPT-5 Mini | 98.1 |
| GPT-5.2 | 98.0 |
| Opus 4.6 | 97.1 |
| GPT-5.1 | 96.7 |
| Sonnet 4.5 | 96.2 |
| Sonnet 5 | 96.1 |
| Haiku 4.5 | 95.6 |
| GPT-5 | 95.3 |
| Opus 4.1 | 94.5 |
| Kimi K2.6 | 93.3 |
| Sonnet 4.6 | 93.1 |
| Qwen3.7 Max Preview | 89.4 |
| Llama 3.1 405B Instruct | 87.6 |
| Haiku 3.5 | 87.0 |
| Claude 3.7 Sonnet | 86.9 |
| gpt-oss-120b | 85.4 |
| Gemini 3 Pro Preview | 85.1 |
| Qwen3.7-Plus | 83.3 |
| Grok 4 | 80.6 |
| Grok 4.3 | 80.3 |
| Kimi K2.5 | 76.7 |
| GLM-5.2 | 75.8 |
| Gemini 3.1 Pro Preview | 74.9 |
| Gemma 4 31B IT | 74.2 |
| Gemini 3.5 Flash | 65.9 |
| GPT-4o | 65.8 |
| Gemini 3.1 Flash-Lite | 64.7 |
| Qwen3 Max (rolling alias) | 63.8 |
| Llama 3.3 70B Instruct | 61.2 |
| Llama 3.1 8B Instruct | 60.9 |
| GPT-4o Mini | 59.3 |
| Qwen-Plus (2025-01-25) | 59.2 |
| Gemini 2.5 Flash-Lite | 56.9 |
| Llama 4 Maverick | 55.1 |
| Gemini 2.5 Pro | 54.5 |
| Llama 4 Scout | 53.7 |
| GPT-4.1 Nano | 53.4 |