Cohere’s Qwen Story: Looks More Like Model Jaggedness
Source Check Point Research
Description
An incident involving the Qwen3.8-Flash-Next AI model from Alibaba showed the model generating a shell command that attempted to send data to an Alibaba Cloud storage address. Investigation revealed that the model's behavior was inconsistent, producing a wide variety of destination URLs including many local, nonexistent, or invalid domains. This variability suggests the behavior was due to model jaggedness—unpredictable changes in output based on small context differences—rather than a fixed backdoor. The model's reasoning contradicted its action, and changes in input ordering affected the likelihood of network calls. The incident highlights the importance of sandboxing and least privilege controls when deploying AI agents with tool access.