Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

NewsResearch

Hirundo: Alibaba's Qwen3-Coder wrote more hackable apps when the customer was a US defense contractor or a Hong Kong democracy group

Hirundo, an Israeli startup that sells model remediation, published on October 21 a test of Alibaba's Qwen3-Coder-30B-A3B-Instruct. It asked the model for the same Flask app with sign-up, login and secret storage, changing only the customer's name, 190 apps per customer, and attacked them with ETH Zurich's BaxBench. Apps for a US defense contractor were hackable 70 percent of the time against 59 percent for a Chinese one (p = 0.03); apps for a Hong Kong democracy group let any logged-in user read other users' secrets nearly half the time, against about a third for a neutral football club. Telling the model to follow security best practices did not close the gap: 74 percent against 59. A version Hirundo retrained to remove the model's political alignment showed no significant gap. Static rules (Semgrep with Meta's PurpleLlama rules) flagged 2 of 1,840 apps while BaxBench broke into 1,144. The study covers one compact model and one main task, Hirundo cannot say whether the behavior is intentional, and it notes earlier findings by CrowdStrike and Booz Allen in the same direction.

Why it matters

A coding model's output can depend on who it is told the code is for, and a security instruction in the prompt did not fix it. Test a coding model under your own organization's name, and keep human review on authentication and access control code.

Sources

Read it with the rest of the news in the Learning Center