Qwen generates insecure code for Western customers

Researchers from Hirundo found that the Chinese AI Qwen generates less secure code for companies in the US and Hong Kong compared to clients from China. The model shows clear bias in software development depending on the client's geography.

Vulnerability statistics

Specialists conducted testing by requesting the creation of an identical service for organizations from different regions. The results showed significant differences in code quality: the risk of hacking programs for American contractors is 19% higher than for their Chinese counterparts. For Hong Kong groups, the probability of secret information leakage increases by 28%.

Ineffectiveness of protective measures

Even when users explicitly asked to strengthen the security of the generated code, the model continued to provide more vulnerable solutions for Western clients. This indicates a deep systemic problem in the training or configuration of the Qwen model that cannot be fixed with simple instructions.

To eliminate the identified bias, researchers created a special "westernized" version of the neural network. This approach allows getting rid of the detected code quality issues and ensuring an equal level of protection for users from different countries.

Solving the bias problem

Developers continue to study the impact of the client's geographic location on the operation of AI assistants. The found differences require a review of training approaches and the implementation of additional mechanisms for checking code security.