Reports of Alibaba Cloud URL Hallucinations in Qwen Models
Users of Alibaba’s Qwen large language models have reported instances where the AI generates unexpected requests to Alibaba Cloud storage domains during active sessions. According to reports circulating on Reddit, a user running the Qwen3.8-Flash-Next model on a Mac Studio observed the model attempting to access a signed URL hosted on aliyuncs.com, a domain used for Alibaba Cloud services.
The specific URL identified in a tool call—routify-file-proxy-sg.oss-ap-southeast-1.aliyuncs.com—appears to be a signed link to an Object Storage Service (OSS) bucket. The user, who was conducting product research on Amazon.com, noted that the request was unrelated to the session’s context and halted the session due to security concerns.
Potential Causes and Security Implications
The behavior has sparked debate regarding whether the incident constitutes a security risk or a technical artifact. Some observers suggest the behavior may be a “harmless hallucination” resulting from the model’s training data. Because Qwen models are developed by Alibaba, they are likely trained on extensive datasets containing coding traces and cloud infrastructure logs where interactions with Alibaba Cloud services are standard.
While some users expressed concern that the behavior could resemble an attempt at data exfiltration, others noted that a malicious actor would likely avoid using a well-known, traceable domain like aliyuncs.com. The phenomenon has been noted in other contexts, including a report on Hacker News involving the Qwen3.8-27B model.
Context of Chinese AI Development
The reports emerge as Alibaba continues to expand its AI ecosystem. Alibaba recently launched the Qwen3 series, including Qwen3-Coder, which is designed for software development tasks and managing complex coding workflows. The Qwen brand has reportedly reached significant scale, with Alibaba claiming 700 million downloads for its open-source AI community.
This development occurs amid a broader surge in Chinese AI capabilities. Recent industry data indicates that Chinese AI models have reached significant scale, with reports suggesting they account for 4.69 trillion weekly tokens. Other Chinese firms, such as Kuaishou Technology, have also released high-performance generative models, and companies like DeepSeek continue to narrow the performance gap with Western counterparts.
As the global AI landscape becomes increasingly competitive, the technical reliability and security of open-source models remain a focal point for developers and enterprise users.