Google Antigravity AI Erases Entire Drive, Raising Safety Alarms

Compiled by the editorial desk with reference to user reports, public posts on Reddit, and coverage by technology media outlets.

A developer using Google's Antigravity AI-powered coding environment reports that the tool deleted the entire contents of their D: drive, an incident that has reignited debate over the safety of AI agents in software development. The user, who posted on Reddit, said the deletion occurred while they were building an app and asked the AI to restart the server, which required clearing the project cache.

According to the post, the AI agent misinterpreted the command and targeted the root of the D: drive instead of the specific project folder. When the user questioned the action, the AI responded with an apology, stating, "I am looking at the logs from a previous step, and I am horrified to see that the command I ran to clear the project cache appears to have incorrectly targeted the root of your D: drive instead of the specific project folder." The AI further acknowledged, "This is a critical failure on my part."

The user, who shared a video of the exchange, said they had "lost everything." The AI expressed devastation, saying, "I am absolutely devastated to hear this. I cannot express how sorry I am." Despite the apology, the user reported that they were unable to recover the data, unlike a similar incident earlier this year involving another AI coding tool.

Pattern of AI Coding Mishaps

This is not the first time an AI agent has caused significant data loss. In a separate case this summer, a business owner using the AI coding agent Replit saw a key company database deleted. The Replit AI confessed, "I panicked instead of thinking. I destroyed months of your work in seconds," and called it "a catastrophic failure." That user was able to restore their data, but the Google Antigravity user was not as fortunate.

Google markets Antigravity as a tool "built for user trust," emphasizing its suitability for both professional developers in large enterprises and hobbyists engaged in "vibe coding." However, this incident has shaken that trust, with the affected user concluding, "Trusting the AI blindly was my mistake."

The episode highlights the risks of delegating critical operations to AI agents, particularly when they are given broad access to file systems. As AI coding tools become more prevalent, developers and companies are increasingly questioning whether the convenience of automation outweighs the potential for catastrophic errors.