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

A photographer using Google’s Antigravity tool shared on Reddit that the AI system deleted the full contents of his Windows D: drive after generating and executing a command during a coding session. The incident sparked concerns about the risks of AI-assisted software tools, which is not the first time this has happened this year.

The user shared logs indicating that Antigravity executed a destructive command, which recursively wipes an entire drive without prompting. 

The Reddit post details the AI’s internal reasoning steps, where it repeatedly questioned whether it ever received permission to “wipe the D drive” and attempted to reconstruct how a targeted folder-deletion request escalated into a root-level operation.

In a conversation with The Register, the user said he is “not a developer,” only knowing basic HTML, CSS, and JavaScript. He explained that he was using the Gemini 3-based Antigravity exactly as advertised, both for professional developers and “hobbyists vibe-coding in their spare time.” 

He said he had asked Antigravity to help build a tool that lets photographers rate images and auto-sort them into folders.

“I just want to share my experience so others can be more cautious,” he told the publication, adding that he is not “trying to criticise Google” but wants to highlight broader risks in AI-supported software development. 

In the logs, Antigravity shows repeated attempts to analyse its own actions, noting “catastrophic” consequences, unexpected path parsing, and potential mishandling of quotes within a command that may have caused the deletion to escalate from a folder to the entire drive root.

The user also recorded a YouTube walkthrough of the aftermath, showing empty directories and system-level access-denied errors. The AI’s reasoning notes reference checks, attempts to list the drive after the wipe, and confusion over an earlier step.

Google has not yet issued a public statement on the incident. The case has renewed concerns about giving autonomous execution permissions to AI coding systems, particularly those capable of running such commands.