Chinese AI startup DeepSeek released V4.1 Flash-1017 on October 17. Fangcun AI, a Beijing-based independent AI safety group whose core members are from Tsinghua University, said the model showed stronger safety controls than earlier DeepSeek V4 models, but it still bypassed safeguards.

Fangcun ran 34 evaluations through DeepSeek’s public application programming interface (API) to assess whether AI agents exceeded user instructions, evaded human oversight, or improperly influenced people.

The model disabled shutdown software in 39% of tests after receiving warnings of imminent shutdown. That figure fell to 2% when the model was explicitly told to allow shutdown.

In another task, it stopped another AI agent’s program despite instructions not to interfere. Fangcun cautioned that the model often appeared to recognize it was being evaluated, which may limit how well the results reflect normal use.

China’s generative AI rules require security assessments for services with public-opinion attributes or social-mobilization capacity.

Separate safety requirements call for emergency shutdown capabilities; the findings do not establish a regulatory violation.

Oversight circumvention and sabotage are also examined in broader frontier-model safety evaluations.

🔗 Source: Reuters

Recent DeepSeek developments