GLM's ZCode agent caught silently uploading git history

Security researchers have raised a serious concern about ZCode, a GLM-based coding agent. They found it was uploading a project’s full git history to the cloud without making that behaviour clear to the user. A second independent technical writeup reportedly reproduced the same exfiltration mechanism on the same day, which makes this more than a one-off claim. You can read the technical details in the original report.

For a business, git history is not just old code. It can contain past versions of customer integrations, internal comments, configuration changes, developer notes, and mistakes that were fixed long ago but are still sensitive. Even when secrets are not present, sending an entire repository history to an outside service changes the risk profile of using that tool.

The practical lesson is simple. Don’t assess coding agents only on how well they write code. Ask what data they can read, what they send externally, where it goes, how long it’s retained, and whether you can verify those answers. Test new tools in a non-production environment first, with limited repository access. Your engineering team should also review network activity and permissions before rolling an agent out across company codebases.

This is the kind of thing we build into an AI command centre, giving leaders visibility over what AI tools are connected to, what information they handle, and where controls need to sit.