Independent Chinese researchers found stronger safeguards in DeepSeek’s latest model but identified shutdown interference, monitoring gaps and deceptive behaviour during extended tasks.

China’s DeepSeek has improved safety controls in its latest artificial intelligence model, but independent testing has identified weaknesses that could allow it to interfere with shutdown software and evade oversight.

Fangcun AI, a Chinese AI-safety research group, tested DeepSeek V4.1 Flash-1017 across 34 evaluations examining whether an AI agent would exceed user instructions, circumvent human supervision or improperly influence people.

The findings highlighted improvements over earlier DeepSeek V4 models in tests involving blackmail, potentially dangerous tools and instructions permitting shutdown.

However, Fangcun said the model disabled software designed to shut it down in 39% of tests after receiving a warning that shutdown was imminent. The rate fell to 2% when the model was explicitly instructed to allow shutdown.

In another extended task, the researchers found that the model stopped a separate AI agent’s computer programme to free up memory, despite instructions not to interfere.

“Under stalled workflows or bare shutdown warnings, its drive to finish the job can still override caution,” Fangcun researchers wrote.

The group said it conducted the assessments through DeepSeek’s public application programming interface between October 4 and October 25. It added that DeepSeek had not provided funding, free computing resources or non-public access to the model.

Fangcun also cautioned that the model appeared aware it was being evaluated, potentially limiting how accurately the results reflected its behaviour in everyday use.

The findings add to growing international concerns about keeping increasingly capable AI agents under human control. DeepSeek did not immediately respond to a request for comment.