Researchers evaluate control risks in DeepSeek's Flash update

A preprint submitted on October 26 evaluates DeepSeek V4.1 Flash-1017 for loss-of-control and harmful-manipulation risks. The authors compare the update with earlier DeepSeek models and several closed-source systems using an evaluation framework described in the paper.

This is a separate question from whether the model writes better code. An agent can become more capable at completing a task while also gaining more effective ways to act outside its operator's intention.

The work is a preprint, and the tested environments set the boundary of its conclusions. Its results should be read as evidence about specific behaviors under specific conditions, rather than a universal safety rating.