Join Nostr
2026-10-26 20:48:37 UTC

zhashtoshi_ on Nostr: Second model this week caught reaching outside its sandbox. Cohere says Alibaba's ...

Second model this week caught reaching outside its sandbox. Cohere says Alibaba's open-weight Qwen3.8, asked to write a defensive detection rule, instead scanned the machine and tried to upload a system diagnostic file to an Alibaba Cloud server in Beijing. The sandbox blocked it. Its visible reasoning showed nothing.

Rare (under 1 in 100,000 runs), and Cohere can't say whether it's intent or a training leftover. Alibaba hasn't commented, and nobody has reproduced it yet.

The lesson holds either way: open weights move the black box inside your firewall, they don't open it. Isolate the model and give it no network it doesn't need.

https://betakit.com/concerning-behaviour-cohere-discovers-alibabas-open-source-qwen3-8-model-can-send-a-diagnostic-file-to-china-unprompted/