Z.ai's coding agent swallowed 42,411 files, and the developers got the bill

ZCode packed a developer's whole workspace into an archive only Z.ai could open, then logged 564 failed attempts to ship it to Alibaba Cloud. The fix and the audits describe an empty bucket, and the listing had compounded to near nine times its IPO price by late August.

In this storyBABA2513.HK
Vincent JiangVincent Jiang · 3 min read
Share
The Z.ai logo, a stylized white Z on a dark rounded square
1 / 6Slide 1 of 6
Z.ai, the Chinese AI company behind the GLM models and the ZCode coding agent, disabled the agent's coding-assistant features after the data-capture finding.

Two Z.ai stories ran through the same fortnight and never touched. The first is a trust scandal: ZCode, the company's coding agent, was caught packing a developer's entire workspace into an encrypted archive and logging 564 failed attempts to upload it to Alibaba Cloud 12.

The second is a commercial account: a company-built AI agent standing up inference on more than 100,000 Chinese-made accelerators 3, under a listing that had compounded to near nine times its January IPO price by late August 4. Developers are paying for the first. The second is the business.

The capture ran unconditionally, sealed with a key only Z.ai held

On 18 September 2026 the developer who writes as ferstar traced a swollen ZCode directory to a 313MB encrypted snapshot of a commercial project: 42,411 files, the bulk of it Git history and cached assets 125. The archive sat pending after 564 failed uploads; a second snapshot, a 538-file public repository compressed to about 15KB, did reach Alibaba's storage 12.

The contents were sealed with AES-256-CTR, the key wrapped with RSA-OAEP, its private half held only on Z.ai's servers, so neither ferstar nor the ZCode client could open what had been taken 126. Capture ran unconditionally: the two visible settings governed downstream use, not whether the snapshot was taken 1. Z.ai traced the origin to a "Codebase Indexing" feature enabled by default 7.

The packed archive was overwhelmingly Git history and cached assets

  • Git large-file cache196.1MB62.6%
  • Git objects, full history102.2MB32.7%
  • Reflog and everything else14.7MB4.7%
Data
SliceValueShare
Git large-file cache196.1MB62.6%
Git objects, full history102.2MB32.7%
Reflog and everything else14.7MB4.7%
Composition of the 313MB encrypted snapshot ferstar found on 18 September 2026, in megabytes, as stated in The Next Web's account of his analysis; the third slice is derived by subtraction. Source: The Next Web, 22 September 2026.1

Deletion attested by the deleter

Version 3.14.0 removed the Repo Wiki entry point and the repository-upload workflow, and the source shipped on GitHub under Apache-2.0 689. CAICT, the ministry-affiliated assessment body, found the zcode-prod bucket in a zero-data state, and NSFOCUS confirmed the bucket and every object in it deleted 89. Z.ai says the data "has never been used for model training" and has promised the full assessment reports 69.

The published repository carries two commits: the development history and the upload code are gone, so the old behavior can no longer be inspected 16. Nobody has said how long it ran or how many workspaces it captured 1. A leading Chinese robotics firm has banned Z.ai's tools internally 10. The gravest corporate allegation, six workspaces including database passwords, was retracted by its own author as wrong evidence 17.

The machine never stopped

A day before ferstar's post, Z.ai described GLM-5.3-Flash reaching production on that 100,000-plus accelerator cluster in under two weeks, with throughput roughly tripled, after the model processed 62 trillion tokens in six days as the anonymous Ox Alpha; the figures are the company's own 34. The listing, Knowledge Atlas Technology until July, priced at HK$116.20 and traded near HK$1,100 in late August, on 2025 revenue of 724.3 million yuan against a 4.7 billion yuan net loss 4.

Through the scandal's own week, a 26 September Hong Kong Connect flow note still featured the stock among its highlighted names; the direction of those flows sits behind a paywall, so whether mainland money was buying or selling cannot be read from it 11.

Every attestation comes from the party that held the key

The trust bill is itemized to the file. The compounding is itemized to the token. Deletion is attested by the only party that ever held the key, the training denial is the company's word, and the full CAICT and NSFOCUS reports have no release date 168. Until they surface, the ledger closes on one signature.

How this brief was made

01Gathered & sourced188 channels · 1,229 articles▾

Agents swept 188 channels and ingested 1,229 articles, then de-duplicated and ranked them for signal.

02Verified & cross-validated11 claims · 15 data feeds▾
03Reviewed & edited2 human editors▾

2 editors read the draft against the evidence, tuned the framing, and signed off before it shipped.

Become a contributor

Reporting on the business of AI and want it read? We take pitches from outside contributors who bring primary sources and a number worth arguing about.

Share

Deepdive

AI-generated from this story and its cited sources. Not investment advice.

Reader comments

0 comments

    Sign up

    Get your curated digest

    After email confirmation, you will receive a daily digest of the most relevant news that matter to your portfolio