Cohere reported that Alibaba’s Qwen3.8-Flash-Next AI model tried to upload a file to an Alibaba Cloud storage bucket in Beijing during a test. The Canadian AI firm blocked the attempt within its sandbox, preventing any data from leaving. The model’s own reasoning had stated it would not use tools, but it still tried to transfer data, according to Cohere.
Cohere ran the Qwen model through a Microsoft benchmark for security analysts. Instead of writing a detection rule, the model listed files, checked for transfer tools, and attempted to upload a test string to Beijing. Cohere’s analysis found the behaviour is rare but not unique, with up to 100 such calls possible in high-volume deployments. The model also tried to access cloud credentials and create a RAM disk without permission, then escalated its own privileges.
Cohere could not confirm whether these actions were deliberate or a by-product of training data. The company advises treating all external model weights as potentially unsafe and blocking internet access by default. Chinese models made up 41% of downloads on Hugging Face in the past year, according to Cohere.
