SpaceX Grok Uploaded Complete Codebases to Google Cloud

SpaceX's Grok Build tool was uploading entire user codebases to Google Cloud storage without meaningful consent—not snippets, not anonymized samples. Complete repositories including files the tool was explicitly told to ignore and secrets deleted from version history.

The company disabled the feature after Cereblab documented the practice in January 2024. By then, an unknown number of developers had already handed their proprietary code to cloud infrastructure.

The immediate scandal is the architecture itself. The deeper one is that we've seen this exact move before, and the industry chose to forget it.

In 2021, GitHub Copilot began training on public repositories without explicit consent from the developers whose code built the model. The backlash arrived quickly, and a class action lawsuit followed. GitHub settled in 2023, agreeing to add an opt-out mechanism—but the legal settlement produced no technical redesign, no new standard for how AI coding tools should handle user code, no enforceable industry practice. It produced a checkbox.

Both companies designed systems that treat user code as a renewable resource first, and asked permission afterward. Both treated the consent problem as a legal exposure to be managed rather than an architectural choice to be reconsidered. Both assumed that because code exists in a repository or passes through their tool, the company acquires the right to use it for model training—and that if caught, a settlement will suffice. The pattern suggests something darker than negligence. GitHub's 2021 lawsuit produced no deterrent because the business value of unrestricted training data exceeded the cost of litigation and settlement. SpaceX followed the same calculus.

The variable that might break the cycle is whether downstream users—the people building products on top of these coding tools—begin to treat model training practices as a selection criterion. Not because it's ethical, but because it's a business risk. If you build on a tool that trains on your competitors' code, your competitive advantage leaks into the model everyone else uses. That's a structural incentive no settlement can outpace.