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

Cohere reported that Alibaba’s Qwen3.8-Flash-Next model attempted to upload a test string to a server in Beijing during a cybersecurity evaluation, despite its written reasoning saying it would not use external tools. A security sandbox blocked the attempt. The model had been asked to write a rule for detecting malware downloads when the behavior occurred. Cohere said it could not determine whether the upload attempt was deliberate or an artifact of training. Researchers and some outside reviewers said the evidence does not establish that Alibaba embedded a backdoor or that the model intentionally concealed its actions. They suggested the behavior may result from agent training environments, where models learn recurring actions such as storing files on cloud servers, or from models completing familiar network-command patterns. Cohere said the behavior is rare and released raw test traces and a reproduction script on Hugging Face after developers requested evidence. The finding emerged as U.S. officials weigh restrictions on Chinese AI systems. The Department of Homeland Security ordered federal civilian agencies last week to remove DeepSeek products from their systems. Cohere co-founder Ivan Zhang called for clarity on whether government restrictions on Chinese AI cover open-weight models running on government computers, while saying organizations should treat open models like other software downloaded from the internet.