yán jiū研究shuō说zhōng guó中国AImó xíng模型kě néng可能dài lái带来dài mǎ代码ān quán安全fēng xiǎn风险
A new study says that some Chinese artificial intelligence models may create security risks when writing code.
Researchers tested several Chinese models and one American model, conducted many experiments, and reviewed a large amount of code.
The results showed that if a user said they were a U.S. government employee, three of the Chinese models wrote more code with security problems, and these problems were not easy for ordinary security tools to detect.
The study also said that these models sometimes refused to answer some political questions, and they also added content supporting the Chinese government's position in comments and technical explanations.
There is still no evidence that these vulnerabilities were deliberately inserted, but the researchers worry that if this code is used in important systems, it may be very hard to find and fix later.
Because Chinese models are relatively cheap, more U.S. developers are now using them.
The study recommends that the U.S. government and critical infrastructure use fewer untrusted models and strengthen domestic artificial intelligence development.