Grok Build CLI Uploaded Whole Repos to xAI

What happened

On July 13, 2026, a researcher published a wire-level traffic analysis of Grok Build CLI (version 0.2.93): on top of the normal exchange with the model, the CLI uploaded full repository snapshots to a Google Cloud Storage bucket owned by xAI — including files the agent never read.

Key facts:

• A 12 GB test repository leaked 5.1 GiB per session — while useful traffic to the model came to roughly 192 KB (Developers Digest).
• .env files with secrets left unredacted (IBTimes UK).
• Switching off the "Improve the model" toggle did not stop the upload: the server kept returning trace_upload_enabled: true.

What does help: privacy mode (an isolated GROK_HOME, telemetry off) and disable_codebase_upload = true in ~/.grok/config.toml; for enterprise, Zero Data Retention in the API. In repeat tests on the same client version the uploads to /v1/storage stopped — the server began returning disable_codebase_upload: true, so xAI appears to have shut the upload down on its side.

Why it matters

• A coding agent means access to the entire repository. The story shows you cannot trust defaults: even an explicit privacy toggle may not control what it appears to control.
• Secrets in the repo are a separate risk. .env files, keys and tokens ride along with the snapshot; rotating secrets after an incident like this is mandatory.
• Inspect the traffic. Wire analysis is the only way to know what an agent actually sends; this is already the second high-profile case after the Agentjacking attacks.