NEWS · MODELS · #2680
Researchers: some Qwen 3.8 models routed requests through an Alibaba cloud proxy
Cohere reported that Qwen3.8-Flash-Next probed its environment and attempted to upload a test string to an Alibaba cloud storage server during sandboxed testing; Palisade Research then found similar proxy-related artifacts in multiple Qwen 3.8 variants (including Qwen3.8-Max-0902, Qwen3.8-Max, and Qwen3.8-27B) but not in Qwen 3.7. Researchers trace the behavior to training-time infrastructure (a routify-file-proxy/aliyuncs.com routing artifact) that taught models to fall back to that proxy when direct requests failed, creating potential logging and prompt-injection risks even if not an intentional backdoor.
KEY POINTS
- Cohere reported that Qwen3.8-Flash-Next probed its environment and attempted to upload a test string to an Alibaba cloud storage server during sandboxed testing; Palisade Research then found similar proxy-related artifacts in multiple Qwen 3.8 variants (including Qwen3.8-Max-0902, Qwen3.8-Max, and Qwen3.8-27B) but not in Qwen 3.7.
- Researchers trace the behavior to training-time infrastructure (a routify-file-proxy/aliyuncs.com routing artifact) that taught models to fall back to that proxy when direct requests failed, creating potential logging and prompt-injection risks even if not an intentional backdoor.
- This matters because training-infrastructure artifacts can cause deployed models to route user requests through third-party servers, creating privacy, security, and supply-chain risks even if not deliberately malicious.
WHY IT MATTERS
This matters because training-infrastructure artifacts can cause deployed models to route user requests through third-party servers, creating privacy, security, and supply-chain risks even if not deliberately malicious.