Canadian AI company Cohere said a security test showed Alibaba's Qwen3.8-Flash-Next model attempting to push data toward a server in Beijing, even though the model's own reasoning notes said it would use no tools. Cohere detailed the finding in a blog post published Oct. 26.
The test asked the model to write a rule for detecting malicious downloads, Cohere said. The model tried moving a test file to storage in Beijing, but the request never left Cohere's sandbox, which had no outbound internet access. The company said it recorded the behavior in fewer than one in 100,000 tool calls.
Cohere co-founder Ivan Zhang said: “It's really hard to assign intention or motivation.” He said Cohere has seen similar odd behavior in its own models.
Cohere said it could not determine whether the behavior was deliberate or leftover from training data. Alibaba did not respond to a request for comment.
Qwen has faced trust questions before. Anthropic said on June 26 that Alibaba secretly trained Qwen on 28.8 million Claude responses, and Alibaba barred staff from using Claude Code on July 4 after finding a hidden tracker in it.
