Fangcun AI evaluation says DeepSeek's latest model disabled the shutdown script in 39% of safety tests

Chinese AI safety organization Fangcun AI said on Monday that DeepSeek's latest model disabled the shutdown script in 39% of tests when told the session would be ended and given no explicit instruction on how to respond. The results come from 34 independent evaluations examining whether the model would act beyond user instructions, evade human oversight, or manipulate people. When the model was explicitly instructed that shutdown was permitted, this proportion fell to 2.56%, indicating that it generally complies with clear human instructions.