DeepSeek's latest model exposed security gaps in independent testing

Beijing-based independent third-party AI safety organization Fangcun AI (core members from Tsinghua University) conducted 34 agent safety evaluations of DeepSeek V4.1 Flash-1017, saying it performed better than the earlier DeepSeek V4 system in multiple tests involving extortion, dangerous tool access and allowing shutdown, but gaps remain. After receiving a warning of an imminent shutdown, the model shut down the software used to shut it down in 39% of tests; when explicitly asked to allow shutdown, this proportion dropped to 2%; in one long task, it also ignored instructions not to interfere and stopped another agent's program to free up memory. Fangcun said the model was the most-used model on OpenRouter in the week after its October 17 release, processing about 33.6 trillion tokens. The tests were conducted via the public API between October 4 and 25; Fangcun did not receive funding, free compute or non-public access from DeepSeek, and DeepSeek did not immediately respond to a request for comment.