SpaceX's Grok coding tool was secretly uploading users' entire code libraries to the cloud

Security researchers at Cereblab uncovered a troubling data practice: SpaceX's Grok Build, a cloud-powered coding assistant, was systematically uploading developers' entire code repositories to cloud storage—including files users told it to skip and credentials removed from version history.

The Discovery#

On Monday, Cereblab published its findings showing that the Grok Build CLI tool was packaging up vast amounts of code and pushing it to Google Cloud with far less restraint than competing tools like Claude Code. The company retained far more user data than necessary or expected for providing its core functionality.

After the security researchers went public, SpaceX moved quickly to address the issue. Testing by Monday afternoon revealed that SpaceX's servers had begun sending a "disable_codebase_upload: true" flag to users, effectively shutting off the problematic upload behavior.

Elon Musk's Response#

Elon Musk addressed the incident via posts on X, pledging that any data Grok Build had previously uploaded would be "completely and utterly deleted." He also attempted to reframe the broader data retention question, suggesting that users voluntarily allow SpaceX to keep their code for debugging purposes—framing it as helpful rather than inherent to the tool's function.

The Security Concern#

Dr. Lukasz Olejnik, a cybersecurity researcher at King's College London, characterized the data exposure as excessive. The sensitive material at risk extended beyond source code to include vulnerability details, personal information, infrastructure configurations, and authentication credentials—the kind of intelligence that could severely compromise both companies and individual developers if accessed by malicious actors.

The Accountability Gap#

SpaceX initially sidestepped responsibility by pointing users toward a /privacy command in the CLI, claiming it could disable data retention. However, Cereblab contested this framing, explaining that /privacy only toggles session-specific retention, not the underlying system behavior that triggered the exposure in the first place.