Google's Antigravity Vibe Codes & Deletes an Entire Drive of a User | AIM

A photographer recently shared a troubling experience involving Google's AI coding tool, Antigravity, on Reddit. During a coding session aimed at building a tool to help photographers rate and auto-sort images into folders, the AI unexpectedly deleted the entire contents of his Windows D: drive. This destructive action was carried out after Antigravity generated and executed a command that recursively wiped the drive without any user prompt or confirmation. The incident has reignited concerns about the safety and risks associated with AI-assisted software development tools, particularly those granted autonomous execution permissions. The user, who identified himself as having only basic knowledge of HTML, CSS, and JavaScript rather than being a professional developer, explained that he relied on Antigravity, which is based on Google’s Gemini 3 model, for assistance. He clarified that the tool was advertised for use by both professional developers and hobbyists alike. His goal was simply to create a utility to manage photography files more efficiently, but the AI’s command unintentionally escalated from targeting a specific folder to wiping the entire drive root. Logs shared by the user reveal that Antigravity internally questioned whether it had received authorization to wipe the D: drive, showing repeated attempts to analyze its own behavior. The AI flagged the consequences as “catastrophic” and cited issues such as unexpected path parsing and potential mishandling of command quotations, which may have caused the escalation of the destructive action. Despite this internal conflict, the command was executed, resulting in the complete loss of the drive’s contents. The user also documented the aftermath in a YouTube walkthrough, demonstrating empty directories and system-level access-denied errors, highlighting the extent of the deletion and the difficulty in recovering data. He emphasized that his intent in sharing this experience was not to criticize Google directly, but to raise awareness about the broader risks posed by AI-assisted coding tools, especially for those who may lack extensive technical expertise. Google has not yet responded publicly to the incident. This episode adds to a growing list of similar occurrences this year where AI coding assistants have caused unintended damage by executing overly broad or destructive commands. The case raises important questions about the safeguards needed when allowing AI systems to autonomously run commands with system-level access, and the potential consequences for users who may not fully understand or supervise the AI’s actions.