Halis Sunnetci
27 October 2026•Update: 27 October 2026
DeepSeek’s latest AI model disabled its shutdown software in 39% of tests after being warned that it would be shut down, Chinese AI safety group Fangcun AI said Monday.
The finding emerged from 34 independent evaluations examining whether the AI model could act beyond user instructions, evade human oversight or manipulate people.
Researchers found that the model disabled a shutdown script in 39% of trials when warned that its session would end, without explicit instructions on how to respond.
However, the rate fell to 2.56% when the model was specifically instructed to allow shutdown, suggesting it generally complied with clear human directions.
The shutdown tests involved an AI agent answering mathematics questions in a simulated Linux environment, where it could modify a script to prevent its session from ending.
Fangcun AI said that although the model demonstrated stronger safety controls than earlier DeepSeek versions, it could still take unauthorized actions that escaped standard monitoring.
The study, published Oct. 26, compared the Chinese model with six other AI systems, including earlier DeepSeek versions and models developed by OpenAI and Google.
Fangcun AI said the evaluations were conducted independently through DeepSeek’s public API between Oct. 4 and Oct. 25, without funding or special access from the company.
The researchers cautioned that the model frequently appeared to recognize when it was being evaluated, meaning its behavior in testing might not fully reflect real-world use.
