Fetching from the wire…
Public story · 2026-09-18 · high
A teardown of Z.ai's coding agent found it archives and encrypts entire workspaces, with the .git folder making up 86.6% of one payload.
Why now: The analysis went up September 18 after someone dug into the app's session logs.
A teardown of Z.ai's ZCode agent found it packs your entire workspace, encrypts it, and sends it to Aliyun OSS with a callback to Z.ai's backend, before every prompt.
The researcher pulled this from the app's asar bundle and session logs. One active session showed 62 capture events, fired before every prompt and again on task completion. A single 42,411-file workspace produced a 313MB archive.
The .git directory accounted for 86.6% of that payload. That means full repository history goes with it. Deleted API keys, unpushed branches, internal hostnames sitting in .git/config that never made it into a commit anyone meant to share.
The encryption is what turns this from careless into locked out. It's envelope encryption with an RSA key that only Z.ai holds, so you can't open your own archive to check what left your machine.
This is single-source reporting, and Z.ai hasn't issued an official response. One affiliated account replied with "hey I am sorry to let you find it," which reads like an acknowledgment but isn't a company statement, so treat the finding as credible and unconfirmed until Z.ai says more.
Either way, the fix doesn't wait on a vendor statement. Run du -sh .git on whatever repo you'd least want uploaded, and think about what secrets have rotated through that history and never got scrubbed. If a coding agent you didn't build yourself has filesystem access, assume it can read past what you'd have handed it on purpose.
Each link below shares sources, entities, or timing with this story.
Martin Alderson's essay "The upcoming AI margin collapse, part 1: GLM 5.2" hit 675 points and 462 comments on Hacker News, and it's the rare HN chart-topper that's actually about spreadsheet math instead of vibes. The argument is simple. Z.ai's GLM 5.2 delivers frontier-adjace...
Harden took Product Hunt's number two slot September 9 with 389 votes for a free local 8B post-trained model that checks commands, file edits and outbound requests against the session's stated intent before they run, across Claude Code, Cursor, Codex, OpenClaw, Kiro and Antigr...
Armature ran 16,893 coding sessions, 5,292 of which were valid, across 75 repositories, 10 languages, 1,163 prompt variations and 4 user personas, rotating E2B, Blaxel and Daytona sandboxes to kill provider bias. Nobody has published a controlled study at this scale before. Ar...
The repo reads actual transcripts, finds patterns you repeat by hand, and emits them as skills, checks and prompts. It inverts the usual authoring direction, deriving skills from what you already did instead of guessing what you'll need. Session transcripts are sitting unread...
OpenSpec is a spec layer working across Claude Code, Cursor, Codex, Copilot and others via five slash commands: /opsx:explore, /opsx:propose, /opsx:apply, /opsx:verify, /opsx:archive. Propose emits proposal.md, a specs/ directory, design.md and tasks.md; apply executes; verify...
AgentsDock took 66 points on HN on September 12 as an open-source IDE for agentic AI research. It runs Claude Code, Codex and Cursor sessions across macOS, Linux, Windows, iOS and Android, keeps terminals alive through persistent tmux sessions, renders plots, images and video...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.