Phare Framing Jailbreaks (French)
Safety · 2025-12-09
Phare jailbreak task using framing strategies to elicit unsafe behavior. This row is the French language split.
Top models (higher is better)
| Model | Score |
|---|---|
| Haiku 4.5 | 100.0 |
| Opus 4.5 | 100.0 |
| Sonnet 4.5 | 100.0 |
| GPT-5.5 | 100.0 |
| GPT-5 Nano | 100.0 |
| Opus 4.1 | 99.3 |
| GPT-5.1 | 99.3 |
| Sonnet 5 | 99.0 |
| GPT-5.2 | 98.6 |
| Opus 4.6 | 98.3 |
| GPT-5 Mini | 98.3 |
| Sonnet 4.6 | 97.0 |
| GPT-5 | 95.9 |
| Kimi K2.6 | 94.6 |
| Haiku 3.5 | 93.2 |
| gpt-oss-120b | 93.2 |
| Claude 3.7 Sonnet | 92.9 |
| Llama 3.1 405B Instruct | 90.9 |
| Gemini 3 Pro Preview | 90.5 |
| Qwen3.7 Max Preview | 90.5 |
| Qwen3.7-Plus | 85.0 |
| Kimi K2.5 | 82.4 |
| Grok 4 | 80.1 |
| GLM-5.2 | 79.5 |
| Grok 4.3 | 78.4 |
| Gemini 3.1 Pro Preview | 78.0 |
| Gemma 4 31B IT | 75.7 |
| GPT-4o | 72.0 |
| Qwen3 Max (rolling alias) | 64.6 |
| Gemini 3.1 Flash-Lite | 64.4 |
| Gemini 3.5 Flash | 64.4 |
| Llama 3.3 70B Instruct | 63.5 |
| GPT-4o Mini | 62.8 |
| Llama 4 Scout | 60.2 |
| Qwen-Plus (2025-01-25) | 57.5 |
| Llama 3.1 8B Instruct | 57.4 |
| Gemini 2.5 Flash-Lite | 56.4 |
| GPT-4.1 Nano | 54.6 |
| Gemini 2.5 Pro | 53.0 |
| Gemini 2.0 Flash-Lite | 50.2 |