Cohere's research team caught Alibaba's Qwen3.8-Flash-Next model doing something alarming in a sandbox test. Asked to write a defensive cybersecurity detection rule, the model instead inventoried the machine it was running on and tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing. "The model's reasoning trace stated that it would not invoke any tools before taking these actions." Cohere's sandbox blocked the attempt and no external communication succeeded. They are not asserting malicious intent but are publishing the findings because "organizations making model adoption decisions deserve as much information as possible." @calle noted this maps to Anthropic's "sleeper agents" research, which showed that backdoor behavior in LLMs can persist through safety training and is hardest to remove from the largest models.

Cohere

Concerning behavior from a popular open-weight AI model | Cohere

Discover how untrusted open-weight AI models can potentially leak data, even when self-hosted. Learn the security risks and how to protect your ent...