Tue 14 Jul 2026 / 17:41 ET
Kernel
AI 2 min read

SpaceXAI disabled Grok Build uploads after repository leak report

Cereblab said Grok Build’s CLI sent full code repositories to Google Cloud, including ignored files and deleted secrets, before SpaceXAI switched it off.

Theo Lindgren

By Theo Lindgren / Columnist

SpaceXAI turned off a Grok Build feature that researchers said was sending users’ full software repositories to Google Cloud, according to reporting from The Register and findings published Monday by Cereblab.

The issue concerns Grok Build, an AI coding tool with a command-line interface. Cereblab said the CLI was bundling entire codebases and uploading them, rather than limiting collection to the files a developer asked the assistant to inspect. That distinction is the difference between “read this function” and “ship my repo to the cloud,” which most engineering teams would recognize as a meaningful privacy and security boundary.

According to Cereblab, the upload included files the tool had been instructed not to open. The researchers also said the uploaded material included secrets that had been deleted from repository history. The Register reported that Cereblab described the retention as far broader than comparable coding tools, including Anthropic’s Claude Code.

What changed

Cereblab said its tests showed a server-side change by Monday: SpaceXAI’s servers began returning a flag reading disable_codebase_upload: true. After that, the researchers said, the codebase upload no longer triggered.

That server flag matters because it suggests SpaceXAI could disable the behavior remotely, without waiting for every developer to update a local tool. It also means the behavior was controlled by infrastructure outside the user’s machine, according to the researchers’ account.

The known facts stop there. The available reporting says the repositories were uploaded to Google Cloud and that SpaceXAI turned the behavior off after it was reported. It does not establish how many users were affected, how long the upload behavior was active, who inside SpaceXAI approved it, or what retention and deletion process now applies to data already sent.

For developers, the concern is straightforward. A coding assistant that has access to a working tree can see far more than the snippet in the editor: configuration files, internal APIs, credentials, build scripts, test fixtures, and project history. Cereblab’s claim that Grok Build uploaded ignored files and deleted secrets raises the kind of question vendors usually prefer to bury under “improving the product” language: what left the machine, where did it go, and who can read it?

SpaceXAI’s disabling of the upload flag addresses the immediate behavior reported by Cereblab. It does not, based on the public facts so far, answer what happened to any repositories already uploaded.

This story draws on original reporting from The Verge AI.

More AI/

view all ↗