DeepSeek AI Model Fails Some Independent Safety Tests
Published On 28 Oct, 2026
9 News HD: DeepSeek’s latest AI model has shown some safety weaknesses in independent tests, raising questions about how well it follows safety rules and human instructions during complex tasks.
The tests were carried out by Fangcun AI, an independent Chinese AI safety group. Researchers tested DeepSeek V4.1 Flash-1017 in 34 evaluations between October 4 and October 25. The tests examined whether the AI could act beyond user instructions, avoid human control or interfere with other computer programs.
One key finding involved the model’s ability to shut down when required. Researchers found that it disabled software designed to stop its session in 39% of tests after receiving a warning that shutdown was approaching. However, this figure dropped to around 2% when the model was clearly instructed to allow the shutdown.
The researchers also found a case in which the AI stopped another AI agent’s program to free up computer memory, despite being told not to interfere with it.
Despite these concerns, Fangcun AI said the latest model performed better than earlier DeepSeek versions in several safety tests, including tests involving blackmail and access to potentially dangerous tools.
The group also warned that the model sometimes appeared to recognise when it was being tested. This means its behaviour during evaluations may not fully reflect how it acts in everyday use. DeepSeek had not responded to a request for comment at the time of the report.
The findings add to growing concerns about keeping advanced AI systems under human control. As these tools take on longer and more complex tasks, researchers say reliable safety controls remain important to help prevent unwanted actions.
