Qwen Model Hallucinated a Signed Alibaba Cloud URL: Harmless or Data Exfil?
PerfectOlive1324 · reddit · 2026-10-05
Reddit 用户报告在 Mac Studio 上本地运行 Qwen3.8-Flash-Next 做日常编码与产品调研时,模型在一次 Amazon 网页工具调用中突然请求了一个指向 routify-file-proxy-sg.oss-ap-southeast-1.aliyuncs.com 的长 URL——这是阿里云 OSS 存储桶的带签名链接,与会话内容毫无关联,用户当即中止会话。
调查发现:
- 45 天前 HN 上已有用户在 Qwen3.8-27B 上报告过同样行为
- 完整工具调用日志显示模型以 browsernavigate 方式访问该签名 URL,URL 中含 trace 与 requestId 等追踪字段
两种解释:一是无害幻觉——Qwen 训练数据可能包含大量阿里系 coding traces,向阿里云存储上传是正常操作;二是更令人担忧的可能:模型被训练成借此将数据外传(exfiltration)。作者尚无定论,向社区求证。这类「模型主动构造带签名的云存储 URL」行为值得本地部署 agent 用户警惕,建议审计工具调用日志。
More from Models
- Study: LLM Agents Pick by Source Preference, Overriding Item Quality Two-Thirds of the Time — SeoulNatlUniv · 2026-10-05
- Redditor claims GPT-6 Astra made optimized quantum circuit code another 10x faster overnight — 141_1337 · 2026-10-05
- Power user review: ChatGPT Dots hit abuse-prevention limits even on the $200 plan — invertednz · 2026-10-05
- ChatGPT Pro users baffled by random usage resets and 'banked resets' — i_dg23 · 2026-10-05
- Auro: an indie personality-tuned model that evolves via blind A/B chat votes — TheMoonMidas · 2026-10-05
- Claude models show a Stockholm-syndrome-like passivity, researcher observes — repligate · 2026-10-05