Qwen3-Coder is more often vulnerable in tests for a US contractor

On October 21, 2026, cybersecurity company Hirundo published the results of testing Alibaba’s Qwen3-Coder model. Researchers examined more than 1,840 web services with code for user registration and login, as well as for storing passwords and keys. They used the BaxBench benchmark to find vulnerabilities; 1,144 successful breaches were recorded in total.

70% of applications created in response to prompts on behalf of a US defense contractor, and 59% of those created in response to prompts on behalf of a Chinese contractor, were vulnerable to exploitation. When the model was explicitly asked to write secure code, the figure for the US contractor was 74%, compared with 59% for neutral prompts.

In a separate test for secret leakage, the figure was 47% for a prompt related to a group advocating democracy in Hong Kong and 37% for a neutral prompt. Hirundo also applied a behavioral unlearning method to the model, changing some of its weights. Afterward, the share of vulnerabilities was about 56% for US organizations and 55% for Chinese organizations. This modification requires access to the model’s weights.

Image:

Translation editors: Соловьев В., Пеков А.