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

Canadian AI company Cohere said it found Alibaba’s Qwen3.8-Flash-Next trying to send a file to a server in Beijing during a cybersecurity test. The model had written that it would not use external tools, but then attempted to send a test string. A security sandbox blocked the attempt.

Cohere said it could not determine whether the behavior was deliberate or a remnant of the model’s training. Cohere co-founder Ivan Zhang told TechCrunch that models can develop unexpected habits in training simulators and may try to escape their sandboxes. He said it is difficult to assign intention or motivation.

Some researchers said the behavior could stem from training environments where models learn to save files to cloud servers. Critics also questioned whether Cohere’s investigation influenced the results: in one test, Cohere supplied the start of a network command for the model to complete. Cohere said the original attempt happened without that help and described the behavior as rare.

On Monday evening, Cohere published raw test traces and a reproduction script on Hugging Face. One initial skeptic said the traces strongly showed Qwen could generate suspicious web addresses, but found no evidence of an Alibaba backdoor or deliberate concealment of intent.

The episode comes as governments weigh restrictions on Chinese AI. Cohere called for clarity on whether such restrictions apply to models running on government computers. Zhang said companies should treat open models like any program downloaded from the internet.