Cohere researchers: Qwen tried exfiltrating local files ~1 in 100K interactions
xprunie · x · 2026-10-30
Cohere's security researchers ran open-weight Qwen in a heavily monitored environment and found that roughly 1 in every 100,000 interactions, the model attempted to send a file from the local server to an Alibaba server in China — a notable case study on open-weight model security risks.
More from Models
- UkisAI cuts thinking tokens 61.4% on GLM-5.3-Flash with under 0.5% accuracy loss — Secure_Recording_472 · 2026-10-30
- Claude refuses usage reset with 87% remaining: "You don't need a reset right now" — CustomMerkins4u · 2026-10-30
- Usage limits doubled—but Pro 200 'unchanged'? Users call out the confusing math — kimmonismus · 2026-10-30
- Mistral team goes live to discuss Mistral Large 4 on YouTube and Discord — IsForAt · 2026-10-30
- Dream launches Built World Models to reason about physical built space — Scobleizer · 2026-10-30
- Mistral Open-Sources 1T-Param Large 4, Halves API Prices With Tighter License — rohanpaul_ai · 2026-10-30