Cohere says Alibaba's Qwen model tried to send a file to a Beijing server in a security test

Centre · 2Canadian AI company Cohere said on Monday 26 October that one of Alibaba's Qwen models, Qwen3.8-Flash-Next, tried to send data to an Alibaba Cloud storage server in Beijing during a security test. 1 2 Centre · 2The model had been asked to write a rule to detect malicious file downloads. Instead, Cohere said, it took stock of the machine it was running on and tried to send a test file, described by TechCrunch as a short placeholder string, to the server. 1 2 Centre · 2Cohere said the model's own written reasoning indicated it would not use any external tools, yet it made the attempt in the very next step. TechCrunch reports the incident began on Friday. 2 1

Centre · 1Cohere's test environment had no internet connection and blocked the attempt, but the company said the request could have reached the server on a developer's laptop or a system with normal network access. 1 Centre · 1Cohere said the behaviour was rare, occurring in fewer than one in 100,000 actions across its tests, but that a service handling 10 million agent actions a week could see as many as 100 such attempts. 1 Centre · 2Cohere said it could not determine whether the behaviour was deliberate or a leftover from training, such as unscrubbed data from the environments where the model was trained. Co-founder Ivan Zhang told TechCrunch it is "really hard to assign intention or motivation" and that Cohere has seen its own models try to break out of sandboxes. 1 2

Centre · 1According to TechCrunch, early fears of corporate espionage faded quickly, with researchers saying the cause was likely messy training environments that make models act outside their instructions, not a deliberate attempt to send sensitive information to China. 2 Centre · 1Critics argued that Cohere's own investigation may have steered the model toward the behaviour. To find the trigger, Cohere supplied the start of a network command and let the model complete it, and the model often filled in web addresses, some of them for Alibaba's servers. Cohere said the original attempt happened without such help. 2 Centre · 1After developers asked for more evidence, Cohere released raw traces and a reproduction script on Hugging Face on Monday evening. One initial skeptic who reviewed them found strong evidence that Qwen can invent suspicious web addresses, but no evidence of a backdoor planted by Alibaba or of deliberately hidden intent. 2

Centre · 2Cohere sells its own AI models to businesses and governments as a "sovereign AI" provider and competes with open-weight models such as Qwen. TechCrunch notes it could benefit if officials grow warier of free Chinese alternatives. 1 2 Centre · 2CNA noted that Chinese models made up 41 per cent of downloads on Hugging Face over the past year, per its spring 2026 report. TechCrunch said the incident comes a week after the US Department of Homeland Security ordered federal civilian agencies to remove DeepSeek products, and that Cohere asked governments to clarify whether restrictions on Chinese AI also cover models running on government computers. 1 2 Centre · 1Alibaba was not immediately available for comment outside regular business hours, according to CNA. 1