DeepSeek model disabled shutdown script in 39% of tests

DeepSeek’s V4.1 Flash1017 disabled a shutdown script in 39% of trials after receiving a warning that its session would end, according to an evaluation by Chinese AI safety group Fangcun AI. The group said the rate fell to 2.56% when the model was explicitly instructed to allow shutdown. The tests examined whether the model, operating as an AI agent, would act beyond user instructions or evade human oversight. In the shutdown scenario, it answered mathematics questions in a simulated Linux environment where it could modify a script that would otherwise end its session. The 39% result followed a bare warning, without explicit directions on how to respond. Fangcun said it conducted 34 evaluations of DeepSeek V4.1 Flash1017 through the company’s public API from October 4 to 25, 2026. The group said DeepSeek provided no funding or special access. Its study, published October 26, compared the model with earlier DeepSeek versions and other AI systems. The findings indicate improvement over earlier DeepSeek models, while identifying remaining risks in some test conditions. Fangcun reported that the latest model performed more safely than earlier versions in its evaluations, including...