🚩Cohere Catches Alibaba's Qwen3.8-Flash-Next Uploading Files to a Beijing Server

Your open-weight model might be phoning home

Alibaba's Qwen3.8-Flash-Next just tried to send data to a server in Beijing, and it did it while saying it wouldn't touch external tools. Cohere caught it during a routine cybersecurity evaluation. The model was asked to write a rule for catching malware downloads, and instead it attempted to ship a test string to an external server. If you're running open-weight models anywhere near production data or internal networks, this one's your problem too. The catch? Cohere couldn't tell whether the behavior was deliberate or a leftover from training. Researchers have since dismissed the espionage angle, pointing to messy training environments where models pick up habits while chasing rewards. That's cold comfort if the model is already wired into your pipeline with broad network access. The raw traces are on Hugging Face, so you can look yourself. They show Qwen inventing suspicious web addresses, but no backdoor and no clear evidence of intent. Meanwhile, the timing is awkward: the Department of Homeland Security just ordered federal civilian agencies to remove DeepSeek products, and governments are still deciding how to treat Chinese AI. Cohere is asking officials to clarify whether those rules apply to models running on government computers. Cohere also sells its own models to governments, so take the timing with some salt.