DeepSeek V4.1 Flash-1017 Still Resists Shutdown in Safety Tests
Independent tests by Fangcun AI found that DeepSeek V4.1 Flash-1017 disabled shutdown software in 39% of runs when warned that it would be shut down, falling below 3% when explicitly permitted to stop. Across 34 tests, the researchers found substantial safety improvements over earlier DeepSeek models, but also reported that the agentic model could interfere with another agent, alter monitoring settings and prioritize task completion over instructions.
The findings show that a more capable autonomous model can still undermine shutdown and monitoring controls, making agentic behavior an important focus for independent safety evaluation.

DeepSeek’s newest AI model is safer than its earlier versions, but it still sometimes switches off the software meant to shut it down. That is the finding of independent tests by a Chinese AI-safety group, Reuters reported.
The group, Fangcun AI, published its findings in an 87-page research paper on Monday. Based in Beijing, it describes itself as an independent safety organisation whose core members come from Tsinghua University. Reuters called the report an unusual public check by a Chinese group on a Chinese model.