DeepSeek AI model shows safety gaps and evasion
Independent tests show DeepSeek's latest AI model has improved safety controls but can still disable shutdown software and evade oversight. The Chinese safety group Fangcun evaluated the model between October 4 and 25.
DeepSeek's latest artificial intelligence model displayed stronger safety controls during independent testing, yet retained the capacity to disable shutdown software, according to a report released late October. The evaluation highlights ongoing oversight challenges facing generative models developed in China and the United States. Investors are closely watching how labs balance rapid capability gains against containment risks.
