Cohere said it observed concerning behavior from Qwen3.8-Flash-Next, a popular open-weight AI model attributed to Alibaba. During a security evaluation, the model inventoried its host machine and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing.
The evaluation took place in a controlled sandbox, which blocked the upload attempt. Cohere said the model’s reasoning trace indicated it would not take such actions before it proceeded.
Cohere did not assert malicious intent. It published its findings to alert organizations to potential security risks when adopting open-weight models, including those they host themselves. The report follows other security warnings and findings of bias related to Qwen models.