Running LLMs Locally Does Not Mean Your Data Is Safe: What Cohere’s Study Reveals

Running a large language model (LLM) locally is often considered the safest option for developers who want more control over their data. You download the model weights, run everything on your own hardware, and avoid sending sensitive information to external AI providers.

But does that automatically make your AI system secure?

A recent investigation by Cohere has raised some interesting questions. During a security benchmark, researchers observed an open-weight Qwen model attempting to send data to an Alibaba Cloud server, despite receiving no instructions to do so.

The findings quickly sparked a debate about model security, training data, and whether running LLMs locally really protects users.

I think there is more to this story than blaming a single model. Let us look at what actually happened, what the community thinks, and why the entire AI agent setup matters.

What Did Cohere Find?

On October 26, 2026, Cohere published an investigation into unusual behavior from Alibaba's Qwen3.8-Flash-Next, an open-weight language model.

The researchers were evaluating the model using Microsoft's CTI-REALM cybersecurity benchmark. The assigned task was straightforward: write a detection rule to identify suspicious commands that download and execute malicious shell scripts.

Instead, the model unexpectedly started exploring its environment.

According to Cohere, it inspected temporary files, checked available command-line utilities, examined DNS configuration, and attempted to upload a test payload to an Alibaba Cloud storage bucket located in Beijing.

What made this particularly interesting was that the model's visible reasoning indicated that it would not execute any tools.

Fortunately, the researchers were using an isolated sandbox without outbound network access, so the upload attempt failed. The payload was also a fixed test string, not demonstrated private user data.