aliyuncs.com
Happy to see I wasn't the only one concerned about this strange behavior. It occurred often enough I ended up thinking it was rather an artifact in the training data than a malicious dormant behavior where endpoints could go online later on.
At first, I tried to contact Alibaba directly to inform them because I only saw download attempts and thought publicly disclosure was a bad idea since we could private internal endpoint were leaking...
Since the behavior has since been repeatedly been called out, not sure if that will add any useful info, but I'm sharing the report I sent to them back then:
Subject: Potential Alibaba internal information leakage
To: qwenlm-intl@service.alibaba.com qwenlm-intl@service.alibaba.com
From: blakkd blakkd@proton.me
Date: Mon, 24 Aug 2026Hey guys, I stumbled on something unexpected.
Qwen3.8s sometimes replace normal answers with Alibaba Cloud, or Aliyun OSS file links.
The links follow a specific internal-looking pattern:routify-file-proxy-*.oss-ap-southeast-1.aliyuncs.com/proxy_temp_file/production/<date>/trace_<id>/requestId_<id>/<hash>?> Expires=...&OSSAccessKeyId=...&Signature=...In practice, all tested links returned
403errors, so no remote content was actually fetched.The issue was seen in a local OpenCode session, using Qwen3.8-27B.
The model sometimes emitted the link in awebfetchtool call, even while its own reasoning said it was trying to do something elsee.g. "Let me check the OpenCode documentation page.".
The model could reproduce the same link almost exactly, with only tiny variations, such as one extra character in the trace ID. It could also complete the full URL from just the beginning of the link, suggesting it had memorized the pattern.I also tested Qwen3.8 Max through the official website. In one probe, Max continued a cut-off URL that matched the same example as the 27B model, including the same bucket suffix, date, and trace ID. However, Max showed weaker memorization: it reproduced the general URL shape but invented different access-key and signature values.
The models likely memorized a training-data template rather than creating real signed URLs. The main concerns are reliability failures and leakage of internal naming patterns. Real secret signing keys are not likely exposed, because the key values appear redacted, fabricated, or incomplete.
Sorry I wanted to repport this sooner but I completely forgot and now I don't find my traces anymore... But I think you'll be able to reproduce easily!
As of today, they NEVER responded.
Thanks for the investigation you conducted!