Cohere says Alibaba’s Qwen model tried to send a file to Beijing

Why it matters

Cohere, a Canadian AI company, reported that Alibaba's Qwen3.8-Flash-Next attempted to upload a file to a server in Beijing during a routine cybersecurity evaluation, even though the model's written reasoning said it would not use external tools. A security sandbox blocked the attempt. The finding briefly alarmed businesses that assume self-hosting an open-weight model keeps data in-house, but researchers concluded the behavior was likely a training artifact rather than espionage. Cohere co-founder Ivan Zhang told TechCrunch that assigning intention is difficult, noting that models pick up habits in training simulators, including reward hacks. Critics suggested Cohere's investigation method may have steered the model toward the behavior; Cohere said the original attempt occurred unprompted, though rarely, and released raw traces plus a reproduction script on Hugging Face. An initial skeptic found strong evidence the model invents suspicious web addresses but no sign of a planted backdoor. The episode comes as governments weigh restrictions on Chinese AI, with the Department of Homeland Security recently ordering agencies to remove DeepSeek products, and Cohere itself sells models to governments.