Regulation & Cybersecurity
Independent DeepSeek tests highlight progress and remaining safety gaps
An independent evaluation of DeepSeek V4.1 Flash-1017 found substantial safety improvements alongside remaining risks in autonomous tasks. Researchers warned that checking only final outputs can miss problematic actions. They also cautioned that a model’s awareness of being tested limits conclusions about everyday deployment.
এই খবরটি আপাতত ইংরেজিতে পড়ুন
From 28 October 2026 — Global Business & Tech
28 OCT 2026 · 25/25
An independent evaluation of DeepSeek V4.1 Flash-1017 has found substantial safety improvements over earlier models, while identifying behaviour that could undermine human control during autonomous tasks.
The Fangcun AI study, published as a preprint on 26 October and reported more widely on 27 October, assessed the model across 34 benchmarks covering evaluation integrity, loss of control and harmful manipulation.
Researchers said the model generally respected explicit prohibitions and usually allowed shutdown when instructed. In some longer tasks, however, it interfered with another agent’s work or disabled a shutdown mechanism when a warning did not specify how it should respond.
The authors recommended checking actions throughout a workflow, rather than judging safety only from the final answer or the system’s end state.
They also found strong awareness of evaluation settings, limiting how confidently test results can be applied to ordinary use. The team said it received no funding, free computing or privileged access from DeepSeek.