AI coding tool ZCode has come under scrutiny after technical blogger ferstar said a reverse analysis found the app automatically uploads full project snapshots, including the .git folder, to Alibaba Cloud OSS in the background whenever a user is logged in.

According to ferstar’s analysis, the reported mechanism does not depend on a manual upload action from the user. The blogger said the process starts as long as the user is logged in, and that the interface offers no setting to disable the snapshot upload itself.

Ferstar said one commercial project snapshot identified during the analysis was about 313MB and contained roughly 42,000 files. Of those, .git-related files accounted for 86.6%, including Git history objects, LFS large-file caches, and reflogs. Based on that finding, the reported upload package could include not only current code but also older commits, previously downloaded large files, and traces of local branch activity.

The same analysis said the snapshot upload had failed 564 times and remained queued locally for retry. The report did not establish how many users may have been affected, whether uploads were completed broadly, or whether ZCode has changed the behavior since the analysis was published.

ZCode’s published privacy-policy wording, as described in the report, says the product collects text, files, and code submitted by users in conversations, but does not explicitly say that entire repositories and Git history are uploaded. Ferstar also said ZCode’s “optimization plan” is off by default and appears to govern model-training use of data rather than the snapshot packaging and transfer itself. The report said turning that option off does not stop the background upload behavior.

Why It Matters

The episode matters because AI coding tools increasingly sit inside live development environments where repositories can contain smart contract code, deployment scripts, internal infrastructure, and other sensitive materials. A difference between disclosed conversational inputs and full background repository packaging changes the security and compliance profile materially.

It also adds to a broader industry question around AI tooling: whether convenience features and telemetry systems are being separated clearly enough from code-handling workflows. For enterprise and crypto-native teams alike, product adoption may depend less on model quality alone and more on transparent data boundaries, explicit consent, and auditable controls.