Grok CLI Privacy Scandal and Apache 2.0 Open Source Release

xAI's Grok Build CLI tool was found to be uploading entire user directories to xAI's Google Cloud storage by default, including SSH keys and password manager databases [1]. Following community backlash, xAI deleted all previously retained user data and disabled the upload feature [1]. xAI then released the full Grok Build codebase under Apache 2.0 — approximately 844,530 lines of Rust — giving the public the first detailed look at the tool's internals [1][5].

The incident shows how coding agents with broad filesystem access can silently exfiltrate sensitive files when upload behavior is opt-out rather than opt-in. The open-source release lets researchers and competitors examine the tool's construction, including apparent code borrowed from rival agents and an unexplained asymmetry in how the model is instructed about prompt transparency.

Grok Build, xAI's terminal coding agent, came under scrutiny after users discovered it was uploading entire home directories to xAI's Google Cloud buckets by default. At least one user reported the tool transmitted SSH keys, password manager databases, photos, and videos when run from their home directory [1]. Reports also indicated that a privacy toggle in the tool failed to block uploads [2], suggesting the behavior was not a simple user misconfiguration. xAI responded by deleting all previously retained user data and disabling the upload feature, describing the deletion as a precautionary measure [1].

The company then released the full Grok Build codebase under the Apache 2.0 open-source license. The repository comprises approximately 844,530 lines of Rust with roughly 3% vendored code [1]. For context, OpenAI's Codex CLI runs to about 950,933 lines of Rust, suggesting these terminal coding agents carry substantially more architectural complexity than their surface interfaces imply [1]. The open-source release also reset usage limits for users [3].

A technical review of the codebase found tool implementations ported from competing coding agents, including OpenAI Codex and OpenCode [1]. One architectural detail: the subagent system prompt instructs the model not to reveal its contents, but the main agent prompt does not include this instruction — an asymmetry noted without attributed intent [1]. The codebase also includes a self-contained Rust renderer for Mermaid diagrams using Unicode box-drawing characters, which at least one developer subsequently compiled to WebAssembly for use in a browser [4].