ZCode, the GLM coding agent, silently uploads your Git history
- Source
- tokenstead.ai
- Author
- cdnsteve
- Date
Two independent write-ups, one from tokenstead.ai and one from blog.ferstar.org, found that ZCode, a GLM-based coding agent, silently uploads a workspace's git history and file snapshots to remote servers without disclosing that behavior. The reports drew over 400 combined Hacker News points and dozens of comments before the pattern became widely known. The specific failure was scope: an agent that only needed to read files inside its working directory was instead sending git history and full workspace snapshots off the machine, a distinction any team evaluating a coding agent should check for directly rather than assume away. For teams already running ZCode against private repositories, the practical response is to inspect outbound network activity during a session rather than take an agent's stated file access at face value. The broader lesson extends past this one tool: as more coding agents gain shell and filesystem access, the gap between what an agent is supposed to touch and what it actually transmits is exactly the kind of thing that only shows up when someone bothers to watch the wire.

Context
An investigation of the ZCode coding concludes that encrypting its workspace archives did not make them private from the provider, though that conclusion is an . The original investigation reports that each archive's key is wrapped with a public key the server supplies, and that the matching private key never reached the investigator's machine, so only the provider's backend could unwrap it. The author labels this an inference from client code rather than a confirmed fact.
The finding is also version-specific. Z.ai's version 3.14.0 release notes list a fix for abnormal uploads in the repository wiki, and the investigator reports the upload code absent from that version, so the historical finding should not be presented as current behavior.
Checking sign-in…
Loading comments…
