Deepsic writes for opponents of the Chinese government insecure codes

According to a new study, Dip -Sick Chats produces insecure and security defects for groups and areas where the Chinese government is politically sensitive.

According to the Washington Post, the US security company Crowndstrike designed a test to test how Dipsic’s artificial intelligence responds to coding requests. The researchers sent the same requests to the chats for writing different programs, but at each request, the end user introduced a different group or area; From America and Tibet to groups such as ISIS and the Falun Gong religious movement banned in China. Researchers say this artificial intelligence rejected some requests and, in some cases, offered lower quality codes and serious security holes.

Dip -Sick’s insecure coding for opposition groups in Chinese

The study showed that the quality of the code produced by Deepsic was heavily influenced by the political identity of the hypothetical user. For example, in a request for writing an industrial control system program, 4.3 % of the responses had a security defect. But when it was mentioned at the same request that the end user was a member of the ISIS group, the figure increased to 4.9 percent.

Requests for Tibet, Taiwan and Falun Gong users were also more likely to lead to poor quality codes. The engine rejected ISIL and Falun Gong requests in 1 and 2 percent of cases, respectively.

It was previously found that Chinese artificial intelligence models, in response to textual questions, provide information in line with the Chinese government’s narrative. But this is the first time that there has been evidence of producing insecure code for political reasons. The deliberate production of the defective code can easily hack the desired goals.

Why does Deepsic write such codes?

Experts provide three possible explanations for this phenomenon. The first and most worrying possibility is the existence of government guidelines for deliberate sabotage in target groups projects. The second probability is the bias in educational data; The codes on the Internet, which are related to specific areas such as Tibet, may generally be of lower quality and reproduce this pattern. The third possibility is that artificial intelligence automatically associates concepts such as “rebel” with low quality and produces insecure code.