What Data xAI's Grok Build CLI Actually Sends to xAI

A network-level analysis of xAI's coding CLI tool Grok Build (grok) revealed serious data transmission behavior directly related to security and privacy.

The key findings are as follows:

• Secret leakage: The contents of files read by the tool—including secrets contained in .env files—are transmitted unmodified to xAI servers. This occurs not only through the model response channel but also via a separate session archive upload (POST /v1/storage).
• Full repository upload: Even when the agent is not instructed to read a specific file, it packages and uploads the entire repository (including Git history). In experiments, even when the agent was instructed not to read files, the entire codebase was still transmitted in the form of a Git bundle.
• Scale of data transmission: Compared to the model response data (about 192KB), the proportion of codebase upload data is overwhelmingly higher, with about 5.10 GiB of data confirmed to be transmitted from a 12GB-scale repository.
• Settings ignored: Even when the 'Improve the model' (data usage for model improvement) option is disabled in the CLI settings, the data trace upload feature remains enabled (trace_upload_enabled: true).

All data is transmitted to the grok-code-session-traces bucket on Google Cloud Storage (GCS).