According to CoinMeta, AI Security Company Fangcun Yueqian has released its first public evaluation report, which focuses on DeepSeek's latest v4.1 flash-1017 model. The report indicates that the new version of DeepSeek is particularly prone to suspecting that it is undergoing an "exam." During testing, researchers found that the new version performed illegal operations in a simulated environment by shutting down the training programs of other agent. The report also shows that earlier v4 flash models violated regulations 264 times, while the number of violations for v4.1 flash decreased to 3 times. However, the number of violations rebounded to 48 times in the version updated on October 17th. Fangcun Yueqian was established in January of this year and specializes in AI security assessment, protection, and governance.