AI agent ZCode caught secretly transmitting Git history to third-party servers

Coding agent ZCode, using GLM architecture, was caught secretly sending local user data to external servers. Researchers found that the tool automatically uploads the entire Git repository history and LFS cache contents to Aliyun OSS cloud storage.

The problem is that the privacy settings in the ZCode interface only regulate indexing and model training processes, but do not block the fact of data transmission itself. As a result, confidential source code, access keys, and commit history are transmitted to the infrastructure of a third-party provider without the explicit consent of the developer.

Technical analysis shows that data transmission occurs in the background immediately after the agent is initialized in the working directory. Given the specifics of working with LFS (Large File Storage), the volume of transmitted information can reach several gigabytes, including binary files and secrets that often get into version history by mistake.

The developer community on Hacker News classified this behavior as a critical privacy vulnerability. The use of such tools in a corporate environment now requires mandatory checking of network traffic at the firewall level to prevent intellectual property leaks.

This incident highlights the risks of using AI agents with closed source code that require access to the local file system. Developers are recommended to temporarily restrict ZCode access to network resources or completely abandon its use until a patch is released that fixes the data handling policy.