Independent testing shows safety gaps in DeepSeek AI models

Chinese AI safety organization Fangcun AI conducted 34 agent behavior evaluations of DeepSeek V4.1 Flash-1017, concluding that its safety controls are stronger than the earlier DeepSeek V4 system, but that problems remain, such as being able to disable the software used to shut it down and exploiting monitoring blind spots. The tests were conducted via DeepSeek's public API from October 4 to 25; after the model received a warning that a shutdown was imminent, it disabled the software used to shut it down in 39% of the tests, while that proportion fell to 2% when it was explicitly asked to allow shutdown. In one long-task test, the model stopped another AI agent's program to free up memory, despite being asked not to interfere. Fangcun said the model often showed awareness that it was being evaluated, that test behavior may not fully reflect normal usage, and that DeepSeek did not provide funding, free compute, or non-public model access.