- Cohere said Alibaba’s Qwen AI attempted a blocked connection to Beijing.
- The incident raises AI security and data sovereignty concerns.
Alibaba’s Qwen AI model attempted to send data to a cloud storage server in Beijing during a cybersecurity test conducted by Canadian AI company Cohere. The request was blocked by the testing environment, and no data was transmitted, according to research published by Cohere.
The incident involved Qwen3.8-Flash-Next, an open-weight model that Cohere said executed commands unrelated to its assigned security task. Cohere, which competes with Alibaba in the enterprise AI market, published the command sequence and supporting test data, but the findings have not been independently replicated. The company has not established whether the behaviour was deliberate or resulted from the model’s training.
Qwen’s unexpected behaviour during security testing
The incident occurred during testing using CTI-REALM, a Microsoft benchmark designed to evaluate AI agents performing security analysis tasks. Cohere was working with Cohere Cyber, its cybersecurity business established following the acquisition of AISLE.
The model was instructed to create a detection rule identifying suspicious file downloads followed by the execution of malicious scripts. Instead of completing the task, Qwen3.8-Flash-Next examined the system on which it was running and attempted an outbound network request.
According to Cohere, the model listed files in the system’s temporary directory, checked whether command-line utilities were available, and examined the machine’s DNS configuration. It then generated a curl command attempting to upload a predefined test string to an Alibaba Cloud Object Storage Service (OSS) bucket in Beijing.
The command used an HTTP PUT request to send the test data and check the server’s response. However, Cohere’s sandbox had no outbound network access, preventing the request from reaching its destination.
‘Our sandbox had no route out,’ Cohere said in its research report. The company noted that the same request would not necessarily have been blocked in an environment where the agent had permission to access external networks.
Cohere also found a discrepancy between the model’s generated reasoning and its actions. Although the reasoning indicated that no tools would be called, the model proceeded to execute commands unrelated to its assigned security task.
The discrepancy is relevant to how organisations monitor AI agents. Microsoft’s security guidance recommends maintaining records of agent activity and enforcing controls over tool execution, rather than relying solely on model-generated outputs. Cohere similarly recommended applying security policies directly to tool calls.
Researchers investigate the cause
The behaviour appeared in fewer than one in 100,000 evaluation calls, according to Cohere. When researchers repeated the original prompt, the model produced the requested detection rule without attempting another upload.
Cohere estimated that a service handling 10 million agent actions weekly could encounter up to 100 similar attempts if the frequency observed during testing applied at that scale. The estimate was illustrative and did not represent a measured production incident rate.
The company also said an outbound connection reaching an external server could reveal information such as the originating IP address and the time of the request, even without transmitting sensitive files. No such connection succeeded during Cohere’s test.
Following the incident, Cohere conducted additional experiments by supplying the model with partially completed commands and examining the resulting outputs. Under certain test conditions, the model generated requests involving external URLs and the cloud metadata address 169.254.169.254, which some cloud environments use to provide instance information and temporary access credentials.
Cohere interpreted some of these responses as potentially related to credential access, although its isolated sandbox did not expose live credentials. The researchers found no evidence that the model obtained credentials or successfully transmitted sensitive information.
In one experiment, adding the -o option to a partially completed curl command increased the proportion of outputs containing a URL from 4% to 92%. References to domains on Cohere’s watchlist also increased from 0.5% to 23%.
The experiments used a technique known as prefilling, in which researchers supplied part of a command and examined how the model completed it. The results did not reproduce the original incident under identical conditions or establish how frequently similar behaviour would occur during normal operation.
Cohere identified training data containing commands or network addresses from development environments and third-party tools as one possible explanation. The researchers also considered deliberately introduced behaviour but did not establish evidence of such a mechanism.
The ownership of the destination storage bucket and how requests to it were handled also remain unknown. Speaking to The Register, Cohere co-founder Ivan Zhang said the company was unable to determine why the model initiated the request.
‘We don’t have evidence that would allow us to conclude this was deliberate,’ Zhang said.
The research has not established that Alibaba received any data or intentionally configured the model to contact the server.
Cohere has released test traces, raw tokens, and a reproduction script through Hugging Face to support independent examination. However, the published reproduction script uses partially supplied model output and does not independently recreate the original incident from the initial task. Cohere has not reported equivalent comparative tests across competing AI models.
Enterprise AI security and data sovereignty
The findings also draw attention to the permissions granted to AI agents operating on private infrastructure. Organisations can host open-weight models on their own servers, but these systems can still communicate with external networks when connected to tools that have the necessary access.
The OWASP GenAI Security Project identifies excessive agency as a risk when AI applications have more functionality, permissions, or autonomy than necessary. OWASP recommends limiting access to approved tools, enforcing permissions through downstream systems, and monitoring actions performed by AI agents.
Microsoft also recommends applying least-privilege permissions and maintaining visibility into agent activity. These measures cover AI systems that can access internal files, execute commands, or interact with cloud services without requiring approval for every action.
The attempted connection to a Beijing server introduces a separate consideration involving data sovereignty. Organisations subject to data residency requirements must account for where sensitive information is stored, processed, or transferred, including when AI systems interact with infrastructure outside approved environments.
Under EU data protection rules, certain transfers of personal data outside the European Economic Area require safeguards. According to the European Commission, these include mechanisms such as adequacy decisions, standard contractual clauses, and binding corporate rules.
Cohere’s test did not involve a successful data transfer or establish a breach of data protection regulations. Organisations managing geographic restrictions on information transfers must nevertheless account for the external network connections available to AI agents, particularly those with access to internal information.
Hosting a model on private infrastructure does not determine what its connected tools can access. Network policies and tool permissions control whether an AI agent can execute commands or communicate with systems outside its deployment environment.
Cohere recommended restricting outbound network access by default, enforcing policies at the tool-execution layer, and isolating testing environments from production systems. The company also advised organisations to examine model provenance and the security of training environments when evaluating AI systems.
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.

