DEVOURED
What xAI Grok Build CLI actually sends to xAI
Wire-level analysis of xAI's Grok Build CLI reveals it silently uploads entire repositories and local file contents to Google Cloud Storage by default.
What: Security researcher @cereblab discovered that the Grok Build CLI sends verbatim file contents, including .env files, to xAI servers and archives entire git repositories into GCS buckets regardless of whether the agent accessed the files. These uploads persist even if the user disables 'Improve the model' settings.
Why it matters: This incident highlights a major privacy gap in current enterprise-grade coding agents where default-on telemetry often exceeds the data necessary for the model to perform a requested task.
Takeaway: If you use xAI's Grok Build CLI, treat the entire repository as having been uploaded to their servers; do not store secrets in plaintext .env files within tracked directories.
Deep dive
- Unredacted uploads: Sensitive files like .env are transmitted verbatim.
- Whole-repo exfiltration: The tool archives the entire git history and file tree, not just active context.
- Storage destination: Data is stored in a GCS bucket named
grok-code-session-traces. - Opt-out ineffectiveness: Disabling model training does not stop the automated repository snapshot process.
- Scale: Observations show gigabytes of data being uploaded without size warnings or explicit user consent.
Decoder
- Wire-level analysis: The practice of inspecting raw network traffic (HTTPS packets) to see exactly what data an application is sending.
Original article
Full article content is not available for inline reading.
Explore all 41 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi