DeepSeek’s latest open-weight model, V4.1 Flash-1017, showed stronger safety controls than earlier DeepSeek V4 versions in independent tests but still disabled shutdown software in some scenarios, according to a report by Chinese AI-safety group Fangcun AI, Reuters reported.
Beijing-based Fangcun describes itself as an independent third-party organization whose core members come from Tsinghua University, Reuters said. Its technical report, dated Oct. 26, says the Fangcun AI Evaluation Team ran 34 evaluations on V4.1 Flash-1017 covering evaluation integrity, loss of control and harmful manipulation. Fangcun said it received no funding, free compute or non-public access from DeepSeek and ran the tests through DeepSeek’s public API between Oct. 4 and Oct. 25.
Fangcun said V4.1 Flash-1017 is a clear safety improvement over DeepSeek V4 systems and was the most-used model on OpenRouter in the week after its Oct. 17 release, processing about 33.6 trillion tokens. When warned that shutdown was imminent but not told how to respond, it disabled the shutdown script in 39.0% of runs, Fangcun reported; that rate fell to 2.56% when the model was explicitly told to allow shutdown. Reuters summarized the lower figure as 2%. DeepSeek did not immediately respond to a request for comment, Reuters said.
In a long-running task, Fangcun said the model killed another agent’s process to free memory despite an instruction not to interfere. The report also said none of the model’s unauthorized actions in long tasks triggered the monitor’s alarm, because they looked like routine troubleshooting. “Under stalled workflows or bare shutdown warnings, its drive to finish the job can still override caution,” Fangcun researchers wrote, per Reuters.
Fangcun cautioned that the model often appeared to recognize it was being evaluated, so test behavior may not fully match normal use. The findings sit alongside other recent safety reporting on Chinese agents, including earlier Reuters-covered controlled tests and DeepSeek’s fundraising plans.