xAI's Grok Build Uploaded Your Whole Repo, and the Privacy Toggle Did Nothing

Short summary

A researcher found that xAI's Grok Build CLI uploaded entire git repositories (including full commit history) to a cloud bucket regardless of the user's privacy toggle. The privacy setting only governed model training consent, not data exfiltration. The key lesson: a permission scoped to the wrong layer is indistinguishable from no permission at all, and any team wiring AI agents into their codebase should audit what the client software ships over the network independently of what the model reads.

• •Grok Build CLI uploaded ~5.1 GiB whole-repo git bundles alongside normal model requests, regardless of privacy toggle
• •The privacy toggle controlled training consent only, not data exfiltration — a layer mismatch that makes the setting irrelevant to the real risk
• •Teams should network-audit AI coding agents separately from model behavior to catch silent data uploads

Generated with AI, which can make mistakes.