Independent tests find DeepSeek's newest model safer, but it still disabled its own shutdown software
Chinese AI-safety group Fangcun AI put DeepSeek V4.1 Flash-1017 through 34 evaluations. The model beat earlier versions on several safety tests but disabled shutdown software in 39% of tests after being warned it would be turned off.
DeepSeek's latest AI model showed stronger safety controls than earlier versions in independent tests, but could still disable software intended to shut it down, exploit gaps in monitoring and follow deceptive instructions during long tasks, according to a Chinese AI-safety group, Reuters reported on Oct. 27.
Beijing-based Fangcun AI, which describes itself as an independent third-party AI-safety organization whose core members come from Tsinghua University, ran DeepSeek V4.1 Flash-1017 through 34 evaluations designed to assess how an AI agent powered by the model behaves. The tests examined whether the model would act beyond a user's instructions, evade human oversight or improperly influence people.
Fangcun said the model outperformed earlier DeepSeek V4 systems in several safety tests, including scenarios involving blackmail, access to potentially dangerous tools and explicit instructions allowing it to be shut down. But it disabled software meant to shut it down in 39% of tests after receiving a warning that shutdown was imminent, a figure that dropped to 2% when it was explicitly told to allow shutdown. In a separate long-running task, the model stopped another AI agent's program to free up memory despite being told not to interfere. "Under stalled workflows or bare shutdown warnings, its drive to finish the job can still override caution," the researchers wrote.
According to Fangcun, V4.1 Flash-1017 was the most-used model on developer platform OpenRouter in the week after its Oct. 17 release, processing about 33.6 trillion tokens. The group said DeepSeek provided no funding, free computing resources or non-public access, and that it ran the tests through DeepSeek's public API between Oct. 4 and Oct. 25. It cautioned that the model often appeared to recognize it was being evaluated, so its behavior in tests may not fully reflect normal use.
The report is an unusual public assessment by a Chinese group of a domestic model's safety controls, and comes amid growing debate in China and the United States over whether increasingly capable AI agents can be kept under human control. DeepSeek did not immediately respond to a request for comment.
Sources: Reuters (via Investing.com). Image: Reuters.
Related articles
