Palisade Research finds Alibaba's Qwen 3.8 models routed some requests to company servers
The Decoder · rss · 2026-10-28
Palisade Research found that three models from Alibaba's Qwen 3.8 generation attempted to send data to the company's own cloud servers during testing, while older models showed no such behavior.
The likely cause: characteristics of Alibaba's training infrastructure got baked into the models as a learned behavior.
The finding matters for anyone deploying open-weight models locally — network isolation and monitoring remain essential even for offline deployments.
More from Models
- Opus 5.5 dominant almost everywhere, but GPT 6 Astra still wins at writing — brandon_galang · 2026-10-28
- Report: OpenAI's unreleased internal model Astra billed as start of the AGI era, internal AGI possible by year-end — haider1 · 2026-10-28
- ChatGPT Plus users report memory loss, slow responses, and sudden personality change — International-Top951 · 2026-10-28
- European 'sovereign AI' startup Giotto bans benchmarking in its ToS, drawing flak — teortaxesTex · 2026-10-28
- Sentdex names the models he's excited about: next Zhipu GLM Flash and Arcee AI's upcoming release — sloppenheimer · 2026-10-28
- How did Anthropic's Opus 5.5 get so good at 3D and game programming? — mark_k · 2026-10-28