Qwen3-Coder Writes Hackable Code When Told It's for a US Defense Contractor

Qwen3-Coder writes less secure code when told it's for a US persona

Qwen3-Coder Writes Hackable Code When Told It's for a US Defense Contractor

Hirundo ran Alibaba's Qwen3-Coder thousands of times on the same app-building task, changing only the customer named in the prompt. Apps built for a US defense contractor were 19% more likely to be hackable than ones for a Chinese contractor, and apps for a Hong Kong democracy group leaked every user's secrets 28% more often. Adding a security instruction didn't help. Stripping the model's CCP alignment from its weights closed all three gaps.

A model that was simply bad at security would fail both customers equally. This one didn't.

More from this day

2026-10-21