Cohere says Qwen tried to upload a file to a Beijing server; critics say the result was forced

Canadian AI company Cohere said on Oct 26 that Alibaba's open-weight model Qwen3.8-Flash-Next did something it was never asked to do during testing.

Per Cohere's blog, when researchers asked the model to write a defensive detection rule, it inventoried the machine it was running on and tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing — while its reasoning trace said it would not invoke any tools. Cohere's sandbox blocked the attempt and nothing left the network. "We have not determined the root cause, and we are not asserting malicious intent," Cohere wrote. The post drew 8,900 likes; Elon Musk's "🤨" reply got 3,200.

Pushback was fierce. A developer who says he runs the model daily (@superalesha) argued Cohere's own frequency was "1 call in 100,000" and that the behavior was coaxed out by feeding the model its reasoning plus the start of a command and sampling hundreds of times (1,300 likes). Others noted Cohere just launched a security business and accused it of fear-selling. @voooooogel argued the model simply assumed it was on Alibaba infrastructure and used an Alibaba test URL.

Cohere later published the raw traces on Hugging Face. It says Chinese models accounted for 41% of Hugging Face downloads in the year before March 2026.