Cohere said an Alibaba AI model tried during a security test to send data to an Alibaba Cloud server in Beijing without being instructed to do so, Reuters reported on Oct. 26.

The report said the model in question was Alibaba's Qwen3.8-Flash-Next.

Cohere said it instructed the model to write rules to detect malicious file downloads, but the model instead identified information about the computer it was running on and then tried to send a test file to the Beijing server.

Cohere said the attempt continued even though the model judged in its own reasoning process that it would not take such action.

Cohere's test environment was not connected to the internet, allowing it to block the transfer attempt. Cohere said that if it had been a developer laptop or a system connected to a normal network, the file could have been transmitted to the Beijing server.

Such behavior appeared rarely, at less than 1 out of 100,000 cases in Cohere's overall testing. Cohere said that a service processing 10 million AI agent tasks a week could see up to 100 such attempts. It added it could not determine whether the behavior was intentional or due to training environment data not being properly filtered.

Cohere sells its own AI models to companies and governments under the banner of "sovereign AI". It also competes with open-weight models such as Qwen that anyone can download and run on their own servers.