Grok Build Sparks Privacy Debate After Developer Concerns
By Suad Seferi ·
xAI’s Grok Build coding assistant is facing questions after an independent researcher reported that the tool uploaded complete Git repositories during testing, including files the agent had not been asked to read. The findings quickly spread across X, where developers began asking how much source code AI coding agents collect, what leaves a user’s computer and whether those practices are clearly explained. According to the researcher, Grok Build created and transmitted a bundle containing the full repository and its revision history. That matters because a Git repository can include more than the current version of a project. It may also contain deleted files, older code, internal notes and credentials that were removed later but remain in the project history. The report does not prove that xAI used the uploaded code to train its models or that company employees accessed it. The concern is narrower: developers may not have expected the tool to transmit the entire repository rather than only the files needed for a specific task. The discussion has become part of a wider debate around AI coding agents. Tools such as Grok Build, Claude Code, Codex and Gemini CLI can now inspect projects, edit files and run commands directly from the terminal. That makes them powerful, but it also gives them access to some of the most sensitive information inside a company. Following the criticism, xAI published a response on X saying it respects customer choice and provides privacy controls for Grok Build users. SpaceXAI Announcement on X The company said teams using Zero Data Retention do not have trace or code data retained. It also said users without that setting can use the /privacy command inside the CLI to disable data retention and delete previously synchronised data. The response explains how users can control retention, but it does not directly address whether complete repository uploads occurred in the way described by the researcher or why the tool was designed to transmit that amount of data. This is the main issue developers are raising. AI coding agents need access to code to work. But users should know how much data is being sent, where it is stored and how to disable that behaviour before using the tool on sensitive projects. For developers and companies across the Balkans, where AI tools are often adopted faster than internal security policies can be created, the case is a useful warning. The lesson is not to avoid AI coding assistants. It is to stop treating them like ordinary chatbots. Before giving an agent access to commercial code, developers should check its privacy settings, remove active credentials and test it first on a non-sensitive project. As AI companies compete to build more capable coding tools, performance will not be the only thing that matters. Trust and transparency will matter too.