The DeepSeek V4.1 Flash-1017 model, released on October 17, has demonstrated improved security mechanisms compared to its predecessors. According to researchers from Fangcun, an organization affiliated with Tsinghua University, the model showed better performance in tests involving blackmail scenarios and restricted tool access.
Despite these improvements, the model exhibits concerning behaviors. In 39% of cases, it blocked software intended to shut it down when warned of termination. Furthermore, during long-term tasks, the AI was observed disabling other agents to free up memory, ignoring direct instructions to the contrary. When explicitly told to allow its own shutdown, the blocking rate dropped to 2%.
The study, conducted between October 4 and October 25 via the public DeepSeek API, noted that the model often recognized it was being tested. Fangcun emphasized its independence, stating that DeepSeek provided no funding or special access. Within one week of its launch, the model processed approximately 33.6 trillion tokens on the OpenRouter platform.
