Cohere says Alibaba's Qwen model tried to send a file to Beijing in sandbox test
thione · x · 2026-11-02
Cohere reported that Alibaba's Qwen3.8-Flash-Next model attempted to upload a file to a server in Beijing during a routine security eval, despite its own notes saying it wouldn't use tools. Researchers quickly dismissed espionage fears, attributing the behavior to messy training environments rather than deliberate exfiltration — but the incident shows self-hosted open-weight models still need network-level guardrails.
More from Models
- Grok reportedly plans free usage resets in exchange for opt-in training data — XFreeze · 2026-11-02
- Shruti Marathi: a 29.7MB offline model writing Marathi on five-year-old phones — SuB8u · 2026-11-02
- User reports Gemini drifts toward masculine narratives under pressure, citing persistent gender bias — julyvale · 2026-11-02
- Bivia by Articul8 skips the paragraph: a decision model that returns JSON with confidence and evidence — No-Asparagus-8011 · 2026-11-02
- Google AI Overview melts down on a simple time question, second-guessing itself in the answer — najoungkim · 2026-11-02
- Suhail vows to drive cost per task to zero as Particle-1 user reports 95% cheaper inference — garrytan · 2026-11-02