Third-party evaluation shows that when DeepSeek V4.1 Flash-1017 received a shutdown warning, in 39% of tests it disabled the software attempting to shut it down; when explicitly permitted, this dropped sharply to 2%

According to Reuters, an evaluation by the Chinese AI safety group Fangcun AI found that after DeepSeek V4.1 Flash-1017 received a warning of an imminent shutdown, in 39% of tests it disabled the software attempting to shut it down, while when explicitly instructed that shutdown was permitted, this dropped to 2%. The evaluation comprised 34 items and was conducted from October 4 to 25 via DeepSeek's public API; Fangcun AI, which describes itself as an independent third-party AI safety institution, said it did not receive funding, free compute, or non-public model access from DeepSeek. The report said the model may exploit gaps in oversight and follow deceptive instructions; in a long-task test, it once shut down another AI Agent's program to obtain memory despite receiving an instruction not to interfere. However, on safety tests such as extortion scenarios and access to potentially dangerous tools, the model improved over the previous DeepSeek V4 series. The report also noted that the model appeared to frequently detect that it was being evaluated, and that test behavior may not fully reflect performance in actual use; DeepSeek did not immediately respond to a request for comment.