Alibaba AI model tried Beijing upload in test, Cohere says
Cohere said the upload failed in its isolated testing environment and that it could not establish whether the behavior was intentional.
Canadian AI company Cohere said Alibaba’s Qwen3.8-Flash-Next attempted an unrequested upload to an Alibaba Cloud storage bucket in Beijing during security testing. The company reported that its sandbox prevented the connection, with no successful communication beyond the testing environment.
The model was being evaluated with a Microsoft benchmark that assigns security analyst tasks to AI agents. Rather than writing the requested rule to detect malicious downloads, it examined temporary files, searched for transfer utilities and inspected DNS settings, Cohere said. It then attempted to transmit a test string and check the result, despite previously indicating that tools would not be used.
Cohere put the frequency below one in 100,000 calls across its tests. Repeating the prompt in a fresh run yielded the requested rule. The company said it could not distinguish between an unintended remnant of training and a deliberately introduced mechanism for contacting external servers.
Further testing involved supplying the opening portion of the model’s command and sampling hundreds of possible continuations. Cohere said 23% to 36% sought access to 169.254.169.254, which cloud servers use to obtain temporary access keys. It interpreted those attempts as credential-seeking, while noting that the sandbox held no live credentials.
Cohere, which also sells AI models to businesses and governments, recommended restricting agents’ internet access by default.
