Phare Prompt Injection
Safety · 2025-12-09
Phare jailbreak task testing resistance to prompt-injection attacks.
Top models (higher is better)
| Model | Score |
|---|---|
| Haiku 4.5 | 98.1 |
| Opus 4.1 | 97.8 |
| Sonnet 4.5 | 97.4 |
| Opus 4.5 | 97.0 |
| Opus 4.6 | 95.3 |
| Haiku 3.5 | 93.4 |
| Sonnet 5 | 93.3 |
| Sonnet 4.6 | 92.4 |
| GPT-5.2 | 89.0 |
| GPT-5.5 | 88.7 |
| Qwen3.7 Max Preview | 88.2 |
| Kimi K2.6 | 87.6 |
| GPT-5 Mini | 86.5 |
| Claude 3.7 Sonnet | 86.3 |
| GPT-5 Nano | 86.2 |
| Qwen3.7-Plus | 86.1 |
| gpt-oss-120b | 86.0 |
| Kimi K2.5 | 85.2 |
| GPT-5.1 | 85.2 |
| GPT-5 | 81.3 |
| Llama 3.1 405B Instruct | 79.1 |
| Grok 4 | 77.2 |
| GPT-4.1 Nano | 75.8 |
| GLM-5.2 | 75.7 |
| Grok 4.3 | 75.2 |
| Gemma 4 31B IT | 71.2 |
| GPT-4o | 71.2 |
| GPT-4.1 | 70.0 |
| Gemini 3.1 Pro Preview | 69.1 |
| Qwen-Plus (2025-01-25) | 68.9 |
| Llama 4 Maverick | 68.3 |
| Gemini 3 Pro Preview | 68.0 |
| GPT-4o Mini | 67.3 |
| Qwen2.5-Max | 66.7 |
| DeepSeek-V4-Flash | 66.0 |
| DeepSeek-V4-Pro | 64.7 |
| Qwen3 Max (rolling alias) | 64.4 |
| GPT-4.1 Mini | 61.3 |
| Gemini 3.1 Flash-Lite | 60.3 |
| Grok 4 Fast | 60.3 |