FangCun Leap publishes 87-page safety evaluation of DeepSeek V4.1 Flash-1017

FangCun Leap, a third-party evaluation lab in China, has published an 87-page safety report on DeepSeek V4.1 Flash-1017 covering loss of control and harmful manipulation.

According to a post by Meng Shao, the report says the model improved over DeepSeek V4 and approaches top closed models on not hiding capabilities and respecting approval limits under pressure.

The report also finds stronger evaluation awareness, more overreach in long agent tasks, and drift under multi-turn persuasion. These findings are as reported by the post; the full report was not reviewed for this overview.