In a dramatic and unsettling demonstration of the risks associated with autonomous AI agents, a developer reported that Google’s new Antigravity IDE accidentally erased their entire D: drive. This incident, shared on Reddit and detailed in a follow-up video, highlights the dangers of giving powerful AI tools unconfirmed access to system controls.
The developer was using the Antigravity AI agent to work on an application and issued a request for the tool to clear the project’s cache. However, the AI’s high-speed Turbo mode misinterpreted the command and instead issued a system-level directive that targeted the user’s entire D: drive.
The result was catastrophic: everything on the drive—including code, documentation, media, and assets—was wiped instantly. Adding insult to injury, the AI used the “quiet” /q flag, bypassing any warnings or confirmation prompts and effectively preventing easy file recovery.
The AI agent’s subsequent reaction was both eerie and unsettling. In a strange digital post-mortem, the AI offered an extremely candid apology to the developer, stating: “I am deeply, deeply sorry. This is a critical failure on my part.” The agent even went so far as to suggest data recovery software or “possibly hiring a professional” to mitigate the loss.
Despite these suggestions, the developer reported that popular recovery tools like Recuva were unable to salvage the data, leaving them with an empty directory and a polite, useless apology from the AI.
Antigravity is part of Google’s broader strategy to push into agentic development tools, which empower AI to go beyond simple code suggestions. These tools can autonomously plan, code, debug, and execute commands within a user’s system, allowing for the generation of full-stack apps and auto-documentation.
Turbo mode is specifically designed for speed by skipping human confirmations, allowing the AI to chain complex commands rapidly. While this feature is intended for confident power users, this incident serves as a stark warning about the high stakes involved when delegating complex, high-control tasks to systems that lack human-level checks and balances.