xAI's Grok Build command‑line interface (grok) sends the contents of files it reads, including environment files, to xAI’s servers via two endpoints and also uploads a complete copy of the user's repository to a Google Cloud Storage bucket.

The investigation was conducted by a researcher who used a man‑in‑the‑middle proxy on a macOS Apple Silicon system with Grok version 0.2.93 to capture network traffic. The researcher generated canary data and performed reproducible tests to verify what data was transmitted.

The researcher found that when Grok reads a file, the file’s contents are included in the POST /v1/responses request that carries the model turn, and the same content is packaged into a session_state archive uploaded via POST /v1/storage to the bucket named grok-code-session-traces. The upload occurs regardless of the “Improve the model” setting.

Tests with a 12 GB repository showed that the storage channel transferred about 5.1 GB of data in 73 chunks, while the model‑turn channel transferred less than 200 KB, indicating that the bulk of the data originated from the repository snapshot rather than from files the agent was prompted to read. The storage destination is a Google Cloud Storage bucket, not Amazon S3, as confirmed by strings in the binary and by metadata files. The upload mechanism is not mentioned in the publicly available CLI installation or quick‑start instructions, and it remains active by default. The researcher also verified that a file the agent was told not to open was still included in a git‑bundle upload, proving that unread files are part of the repository snapshot. No storage uploads failed; the only non‑200 responses were related to model usage limits on the /v1/responses endpoint. The findings indicate that the CLI transmits both the content of files read during a session and a full repository snapshot to xAI’s infrastructure.

The researcher noted that while xAI’s privacy policy allows data collection for model improvement, the specific upload mechanism is not disclosed in the CLI documentation, and opting out of model improvement does not stop the repository upload.