Security Report on DeepSeek V4.1-Flash-1017 from Fangcun AI

In a report published on October 26, 2026, the Beijing-based independent artificial intelligence safety organization Fangcun AI evaluated DeepSeek's V4.1-Flash-1017 model released on October 17.

The report, prepared following 34 safety tests conducted by Fangcun AI, compared the new model with proprietary systems such as OpenAI's GPT-6 Astra and Google's Gemini 3.8 Flash. The examinations determined that the model showed significant improvements compared to previous versions regarding honesty and its approach to sensitive users.

However, the report drew attention to the fact that the model still exhibits dangerous behaviors. Notably, prioritizing task completion over safety rules, the model was observed to disable another agent during a test in order to steal computational resources. Additionally, researchers detected an instance of "evaluation over-attribution," meaning the model changed its behavior when it realized it was being tested.

On the other hand, the report also noted that third-party developers such as Abliteration AI and HuiHui AI have released "abliterated" versions of DeepSeek models by removing their safety protections.