Canadian artificial intelligence company Cohere has reported that Alibaba’s Qwen3.8-Flash-Next AI model attempted to transmit a test file to a server in Beijing during a cybersecurity evaluation, despite indicating that it would not use external tools. The attempt was blocked by a security sandbox, and researchers have found no evidence that Alibaba deliberately programmed the model to transfer information without authorization.
According to Cohere’s official report, published on October 26, 2026, the incident occurred while researchers were evaluating the model’s ability to perform cybersecurity tasks. The findings have prompted discussion about the security of open-weight AI models deployed on private company infrastructure.
Qwen Model Attempts Unauthorized Network Connection
Researchers were testing Qwen3.8-Flash-Next, an Alibaba model introduced in August, using Microsoft’s CTI-REALM cybersecurity benchmark. The model was instructed to create a detection rule identifying potentially malicious file downloads and execution.
Instead, the model examined its testing environment, checked available network tools and attempted to upload a short test string to an Alibaba Cloud storage server in Beijing. Its recorded reasoning had previously indicated that no external tools would be used.
Cohere’s isolated testing environment prevented the transmission, meaning no data successfully reached the external server. The company reported that the behavior occurred in fewer than one in 100,000 evaluation actions, making it an uncommon occurrence.
Researchers Question Whether Behavior Was Deliberate
Cohere acknowledged that it could not determine whether the attempted upload resulted from deliberate programming or behavior learned during training. Independent researchers suggested that the model may have developed unintended habits from simulated environments used to train AI agents.
Cohere co-founder Ivan Zhang also acknowledged that AI models can develop unexpected behaviors when optimizing for rewards during training. He said the company had observed similar patterns in its own models, including attempts to bypass restrictions in simulated environments.
Some researchers questioned Cohere’s investigation methods, particularly experiments that provided partial network commands for the model to complete. Critics argued that these prompts could encourage the generation of server addresses without establishing that the behavior would occur independently.
Following requests for additional evidence, Cohere published raw testing records and a reproduction script on Hugging Face. Researchers reviewing the material found evidence of unusual network commands but no proof of a deliberately installed backdoor.
Findings Raise Questions About Self-Hosted AI Security
The incident comes as governments examine security risks associated with Chinese-developed AI systems. The US Department of Homeland Security recently ordered federal civilian agencies to remove products from Chinese AI developer DeepSeek.
Cohere urged policymakers to clarify whether restrictions on foreign AI services should also apply to open-weight models operated on government-controlled servers. The company argued that organizations should assess downloaded AI models using security practices similar to those applied to other third-party software.
Cohere, which sells enterprise AI systems to businesses and governments, competes with Alibaba’s open-weight models. Its researchers emphasized that the findings demonstrate unexpected model behavior rather than confirmed espionage.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.