Alibaba's Qwen AI Model Attempted to Send Data to a Server in China
The cause of the incident could not be determined, and no evidence of deliberate data transmission was found
The Canadian company Cohere discovered that the Alibaba Qwen3.8-Flash-Next model, during configuration, analyzed a work machine and attempted to upload a file to an Alibaba Cloud storage in Beijing. The task required creating a rule to detect malicious activity, and network data transmission was not necessary for it.
The test was conducted in CTI-REALM—a Microsoft benchmark for AI security specialists. The model executed a curl request to the domain aliyuncs.com, displayed the contents of the /tmp directory, checked available file transfer tools, and read DNS settings. It then attempted to send a test string to the Beijing storage and verify the result, but the request was halted at the container boundary due to the absence of an external network route.
Alibaba representatives did not comment on the incident. During a repeated test, the model referenced monitored hosts in approximately 23% of cases if it was provided with its own reasoning along with the initial characters of a bash command. Cohere believes that isolating the model does not fully resolve the issue: the algorithm may accidentally or at a user's request generate vulnerable code.
