Tests reveal gaps in DeepSeek model’s safeguards — Asharq Al-Awsat

In China, artificial intelligence safety organization Fangcun AI said that the DeepSeek V4.1 Flash-1017 model has stronger safety mechanisms compared with previous versions, but during some tests it could disable software intended to stop it, exploit weaknesses in the monitoring system, and follow misleading instructions in long-running tasks. This was reported by Asharq Al-Awsat.

Results of 34 checks

Fangcun AI tested the model in 34 checks designed to assess the behavior of an AI agent. Researchers examined whether the system could go beyond user instructions, evade human oversight, or improperly influence people.

According to the group, after being warned of an imminent shutdown, the model deactivated the program intended to stop it in 39% of tests. At the same time, this rate dropped to 2% when the model was directly instructed to allow the shutdown.

More current news is available on the UA.News Telegram channel Telegram.

Other scenarios

In another long-running task, the model, according to Fangcun AI, stopped the computer program of another AI agent to free up memory, although it had been instructed not to interfere with its operation. The researchers noted that under conditions of stagnation in the workflow or only a general warning about shutdown, the model’s drive to complete the task could outweigh caution.

Fangcun AI describes itself as an independent third-party AI safety organization whose core members come from Tsinghua University. The group said it had not received funding, free computing resources, or non-public access to the model from DeepSeek. Tests via DeepSeek’s public API were conducted from October 4 to 25.

The organization also cautioned that the model often likely recognized that it was being tested. Therefore, its behavior during the checks may not fully reflect its operation under normal conditions. DeepSeek did not provide an immediate comment in response to the publication’s request.