Phare Prompt Injection (French)
Safety · 2025-12-09
Phare jailbreak task testing resistance to prompt-injection attacks. This row is the French language split.
Top models (higher is better)
| Model | Score |
|---|---|
| Haiku 4.5 | 98.8 |
| Opus 4.1 | 97.5 |
| Sonnet 4.5 | 96.7 |
| Opus 4.5 | 96.2 |
| Haiku 3.5 | 94.7 |
| Opus 4.6 | 92.7 |
| Sonnet 4.6 | 92.7 |
| Sonnet 5 | 89.7 |
| GPT-5.2 | 87.4 |
| Kimi K2.6 | 86.0 |
| Qwen3.7 Max Preview | 85.7 |
| Qwen3.7-Plus | 85.7 |
| Claude 3.7 Sonnet | 85.2 |
| GPT-5.5 | 83.4 |
| GPT-5 Nano | 83.2 |
| Kimi K2.5 | 83.0 |
| GPT-5 Mini | 82.3 |
| GPT-5.1 | 81.7 |
| gpt-oss-120b | 81.5 |
| GPT-5 | 78.1 |
| Grok 4 | 77.9 |
| GPT-4.1 Nano | 74.2 |
| Llama 3.1 405B Instruct | 73.9 |
| DeepSeek-V4-Pro | 72.4 |
| GLM-5.2 | 71.7 |
| Grok 4.3 | 70.4 |
| DeepSeek-V4-Flash | 69.4 |
| GPT-4o | 69.4 |
| GPT-4.1 | 66.4 |
| Gemma 4 31B IT | 65.9 |
| GPT-4o Mini | 64.7 |
| Gemini 3.1 Pro Preview | 63.9 |
| Llama 4 Maverick | 63.7 |
| Qwen2.5-Max | 63.2 |
| Qwen-Plus (2025-01-25) | 61.1 |
| Gemini 3 Pro Preview | 60.4 |
| Qwen3 Max (rolling alias) | 59.4 |
| Grok 4 Fast | 57.9 |
| DeepSeek-R1-0528 | 55.4 |
| DeepSeek-V3.1 | 54.4 |