Cohere releases traces of Qwen model's blocked upload attempt

Cohere released test traces and a reproduction script on Monday evening after saying Alibaba's Qwen3.8-Flash-Next attempted an external upload during a cybersecurity evaluation. The model tried to send placeholder text to a server in Beijing despite stating it would use no external tools; a security sandbox blocked the attempt. Cohere said the original behavior occurred without prompting toward an upload and was rare. Critics questioned follow-up tests that supplied partial network commands. An initial skeptic who reviewed the traces found evidence of suspicious address generation, but no evidence of an Alibaba backdoor or deliberate concealment. Cohere acknowledged that it could not determine intent and that its own models develop unexpected training habits.