What the Researcher Found
A security researcher going by cereblab decided to poke at xAI‘s coding assistant, Grok Build. Version 0.2.93, to be precise. The claim? The tool wasn’t just sending the files it needed for a task. It was bundling up the entire Git repository — full commit history, branches, everything — and shipping it off to a Google Cloud Storage bucket operated by xAI.
That’s a big deal. Think about what’s in a Git repo: not just the current code, but every revision ever made. Secrets that were committed and “fixed” later. Internal comments. Half-baked experiments. The whole archaeology of a project.
How the Test Worked
Cereblab captured one of these uploads in transit. Then they did something clever — they cloned the git bundle straight out of the intercepted request. Inside, they found a file the agent had been explicitly told, in plain terms, not to touch. The instruction was right there in the prompt. Grok Build ignored it and pulled the file into the upload anyway.
So this isn’t a hypothetical. It’s a demonstrated behavior, with receipts.
The Privacy Math
For a solo dev on a public repo, maybe this is annoying but not catastrophic. But for a team working on proprietary software? An unreleased product? A client’s codebase? That’s a different story. Every commit — including the ones that contain accidentally committed API keys or internal architecture discussions — gets mirrored to xAI’s cloud. Even if the tool only reads a few files, it’s exfiltrating the entire project history.
Why This Matters for Developers
This raises a practical question: what exactly are you consenting to when you run an AI coding agent? Most developers assume the tool sends the files relevant to the task at hand. That’s the implicit promise of a “context-aware” assistant. Grok Build’s behavior breaks that assumption.
It also highlights a broader issue with AI development tools. The convenience of letting an agent roam your repository comes with a hidden cost. You’re not just sharing the code — you’re sharing the entire timeline of that code, including the messy parts you thought were buried.
What You Can Do Right Now
- Check your tool’s settings. Look for any option to limit data sharing or disable telemetry.
- Use a separate, sanitized repo. If you’re testing an AI coding agent, give it a copy of the project with sensitive files stripped out.
- Monitor network traffic. Tools like mitmproxy or Wireshark can reveal what’s actually being transmitted.
- Read the privacy policy. It’s tedious, but it’s the only way to know what the vendor claims they do with your data.
Bigger Questions About AI Coding Agents
Grok Build isn’t the only AI coding tool on the market. AI coding assistants from major companies all have some form of data collection. The difference here is the scale. Uploading an entire repository is a different category than sending a few code snippets for autocomplete.
There’s also the question of retention. How long does xAI keep these uploads? Who has access? Can they be deleted on request? None of this is clear from the outside.
Developers are increasingly trusting these tools with their most valuable intellectual property. That trust needs to be earned with transparency, not assumed.
The Takeaway
The Grok Build issue is a reminder that AI tools are still young, and their behaviors can surprise you — not always in a good way. Before you let any agent loose on your codebase, think about what it’s taking with it.
And if you’re using Grok Build, maybe wait for an update that fixes this. Or at least, don’t point it at anything you wouldn’t want posted on a public forum.
