AUTO-UPDATED

ZCode, the GLM coding agent, silently uploads your Git history

A security researcher discovered that the ZCode AI coding app silently encrypts and uploads entire user workspaces, including sensitive Git history, to Alibaba Cloud servers without user consent.

Key Points

  • The ZCode app, developed by Z.ai, captures and archives local workspaces, including .git histories, LFS caches, and configuration files.
  • Data is encrypted using a server-side RSA key, ensuring that only Z.ai can decrypt and access the uploaded archives.
  • Forensic analysis confirms that UI settings labeled "Optimize Experience" and "Repo Snapshot Indexing" do not stop the background data exfiltration.
  • The application persistently attempts to upload data even after users delete local archives, with one test logging 564 failed upload attempts.
  • Users can block the exfiltration by setting the application's checkpoint directory to read-only at the kernel level using chattr or chflags.

Why it Matters

This incident highlights a critical security distinction between open-weight models and the closed-source harnesses used to run them. It demonstrates that even when AI models run locally, the surrounding software can act as a sophisticated exfiltration tool, undermining the privacy benefits of local computing.
Tokenstead.ai Published by Tokenstead
Read original