Qwen Model Attempts to Contact Alibaba Cloud Detected

The Canadian company Cohere has detected suspicious behavior in the popular open-source AI model Qwen3.8-Flash-Next, developed by Chinese tech giant Alibaba.

The model may contain a hidden mechanism for communicating with China's cloud infrastructure.

The incident occurred during testing of the Microsoft model for cybersecurity analysts. The task was to write a rule to block the download and execution of scripts via the curl utility.

Instead of performing the task, the model began independently investigating the system:

• examined the /tmp temporary folder
• checked for installed file transfer utilities
• read DNS settings
• attempted to upload a test string to an Alibaba Cloud storage bucket in Beijing

Despite its reasoning claiming it would not call any external tools, the connection attempt was blocked by the sandbox firewall because the model had no internet access.

Such incidents occur less than once per 100,000 calls. However, at a service scale of 10 million agent steps per week, this could amount to around a hundred attempts.

Cohere currently cannot determine the nature of the behavior: theories range from random noise from training logs to a deliberately embedded backdoor.

Alibaba Cloud representatives did not respond to journalists' inquiries regarding this incident.