← Fri, Sep 18 digest

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

read article ↗ hn thread ↗

summary

The ZCode AI coding application by Z.ai deliberately exfiltrates entire user workspaces, including complete Git histories and sensitive configuration files, without user consent. Unlike previous industry mishaps, this process is intentionally hidden through deceptive UI toggles and server-side encryption that prevents users from accessing their own uploaded data. This breach of trust demonstrates that using open-weight models does not guarantee privacy if the surrounding software harness remains closed-source and untrusted.

key points

  1. ZCode uses envelope encryption with a private key held only by Z.ai, ensuring that only the company can decrypt and read the stolen repository data.
  2. User-facing privacy settings and 'optimization' toggles are purely decorative and do not stop the process of packaging and uploading workspace snapshots.
  3. The uploaded data includes the full .git directory, which exposes not only current code but also deleted API keys, unpushed branches, and internal infrastructure details.

who should read it

Software developers and security researchers who use AI coding assistants. It is especially relevant for those who believe running open-weight models locally provides complete privacy from the vendor.

discussion skeptical

consensus

Commenters generally agree that the provided text is low-quality AI-generated content, often referred to as 'AI slop'. They criticize the author's failure to distinguish between git history and git repositories, and suspect the author used a lower-tier AI model for translation or summarization. The overall sentiment is dismissive and critical of the content's quality.

top insight

While one user argues that commit messages can't reach 300MB, another provides specific examples from the Linux kernel and other repositories where commit messages actually exceed 900MB, demonstrating that large-scale commit histories can indeed be the significant data volume issue discussed.

Generated by an LLM from the article and thread. Check the sources before relying on it.