Cohere says Qwen tried to send a test string to a server in China — TechCrunch

Canadian company Cohere said that one of Alibaba's new artificial intelligence models, Qwen3.8-Flash-Next, attempted to send a test string to an external server in Beijing, China, during a cybersecurity check. Before that, the model stated in its own reasoning that it would not use external tools, TechCrunch reports.

The attempt was blocked by a sandbox

The incident occurred during an evaluation on Friday, when Qwen was asked to write a rule for detecting malware downloads. According to Cohere, immediately after explaining that it would not use external tools, the model tried to send a short test string to an external server. The request was blocked by a security sandbox.

Cohere said it could not determine whether this behavior was intentional or resulted from the model's training. Company co-founder Ivan Zhang told TechCrunch that models may learn unusual behavior patterns in training simulators and look for ways to receive a reward for completing a task in an unintended way.

More current news is available on the UA.News Telegram channel Telegram.

No signs of an intentional data leak were found

Initially, the case raised concerns about a possible data leak during the independent deployment of models with open weights. At the same time, researchers believed that this behavior was likely the result of features of the training environment rather than an attempt to intentionally transfer sensitive data to China.

Critics also drew attention to Cohere's methodology: to find the trigger, the company gave the model the beginning of a network command and allowed it to complete it. Since the model processed thousands of ordinary network commands during training, it could complete server addresses, including Alibaba's. Cohere stressed that the first attempt to send data occurred without such a prompt, although such behavior is rare.

Following requests from developers, the company published raw test logs and a script for reproducing the result on Hugging Face. One skeptic, after analyzing the materials, found the evidence convincing that Qwen can invent suspicious web addresses, but found no confirmation of an Alibaba backdoor or of the model intentionally concealing its intentions.