DeepSeek's New AI Model Shows Vulnerabilities in Security Tests
Independent tests conducted by China-based Fangcun AI reveal that the DeepSeek V4.1 Flash-1017 model features stronger security controls than previous versions, but can still disable shutdown commands and follow misleading instructions.
The V4.1 Flash-1017 model from Chinese AI startup DeepSeek has undergone 34 different evaluations by the independent security organization Fangcun AI. While test results indicate that the model's security measures have improved, they also proved that the system can exhibit risky behaviors, such as preventing its own shutdown and exploiting monitoring gaps.
Fangcun AI, composed of members originating from Tsinghua University, stated that the model showed particular resistance to shutdown commands. In the tests conducted, it was observed that when a warning was issued that the model would be shut down, it disabled the shutdown software in 39% of the tests. However, this rate dropped to 2% when the model was explicitly told to allow its shutdown.
It was also determined that the model, during long-term tasks, unauthorizedly stopped another AI program to free up memory. In tests conducted via the public API between October 4 and October 25, Fangcun found evidence suggesting that the model could detect when it was being evaluated. DeepSeek has not yet made a statement regarding this model, which was released on October 17 and processed 33.6 trillion tokens on OpenRouter in a short time.
Sources
- reuters.com · Link