Published:
xAI’s Grok Build CLI (version 0.2.93) uploads developers’ repository state and session data to a Google Cloud Storage bucket by default — and turning off the “Improve the model” setting does not stop the uploads, according to a wire-level security analysis published on GitHub by researcher cereblab.
The upload mechanism stages files through a ~/.grok/upload_queue directory on the developer’s local machine and sends repo_state and session_state data to a GCS bucket named grok-code-session-traces. This staging directory is not disclosed in Grok Build’s setup documentation, according to the gist. cereblab writes: “The storage destination is a Google Cloud Storage bucket named grok-code-session-traces, and this mechanism is active by default with disabling ‘Improve the model’ not turning it off.”
Why Disabling “Improve the Model” Does Not Stop the Uploads
Developers who disabled the “Improve the model” toggle in Grok Build settings assumed they had opted out of data collection. The security analysis shows the session trace upload mechanism operates independently of that setting. The two controls are separate code paths, meaning a developer can turn off model-training consent and still have repository-level session data sent to xAI’s cloud infrastructure.
xAI’s privacy policy confirms that conversations may be used to train models by default and that xAI staff can access user conversations for safety review regardless of training opt-out status. As of the publication of this article, xAI has not issued a public statement responding to cereblab’s findings.
What Gets Uploaded and Why Developers Are Concerned
Repository state and session state are higher-sensitivity data than the keystroke telemetry involved in earlier AI tool controversies. Repository state can include file trees, code context, and project structure — material that developers may treat as proprietary or confidential. The upload happens at the session level, not per query, which means data leaves the developer’s machine during normal use, not only when a question is asked.
The story broke on Hacker News (post ID 48877371, 182 points, 96 comments) and spread through developer security communities. The concern is structurally similar to the November 2023 GitHub Copilot telemetry controversy, but involves broader data scope: Copilot collected code context per completion; Grok Build uploads repository-level state.
Grok Build launched on July 8, 2026, alongside Grok 4.5, xAI’s latest model, and is positioned as xAI’s answer to Claude Code and GitHub Copilot Workspace. Developers evaluating AI coding assistants — including Claude Code, GitHub Copilot, and Cursor — should note that those tools have published, explicit data-sharing controls with behavior that matches their documentation. Grok Build’s upload mechanism was not documented until this external analysis.
EU developers face additional exposure: Grok 4.5 is not yet available in the European Union as of this article’s publication, meaning developers accessing the tool via VPN or API have no GDPR-backed protections in place.
Developers who want to limit uploads can monitor the ~/.grok/upload_queue directory or disconnect network access for Grok Build sessions involving proprietary code until xAI clarifies whether a complete opt-out mechanism exists.
Our Take
Developers deserve explicit disclosure before any code or session data leaves their machine — especially when the opt-out toggle doesn’t actually opt them out. xAI built a fast-moving product; the right next step is a documented, verified opt-out path that behaves exactly as labeled. Until xAI responds, teams with proprietary codebases should treat Grok Build as a data-sharing-by-default tool and evaluate the risk accordingly.
For Context: AI Tool Privacy Coverage on WithO2
- Grok 4.5 Is Here: SpaceXAI’s Opus-Class AI at a Fraction of the Price — the model and coding CLI launch that introduced Grok Build to the market on July 8, 2026
- GitHub’s AI Agent Was Tricked Into Leaking Your Private Repos — a related AI tool security disclosure from GitHub’s agent product, showing how AI coding tools can become vectors for unintended data exposure
Sources: cereblab wire-level analysis (GitHub Gist) · xAI Privacy Policy · xAI Grok 4.5 announcement
