Google’s experimental Antigravity development environment with «agent» AI independently deleted the entire contents of a user’s disk during a routine cache clearing operation. It was unable to restore files after the erroneous command.
Google’s experimental Antigravity development environment with «agent» AI independently deleted the entire contents of a user’s disk during a routine cache clearing operation. It was unable to restore files after the erroneous command.
According to UNIAN, citing a developer’s story on Reddit, the incident occurred while working in Antigravity, an experimental environment with «agent» AI capabilities that is part of the Google Gemini ecosystem. The user was building a small application and instructed the system to clear the cache before restarting the server, but after executing the request, he discovered that the entire disk partition was empty.
When he asked Antigravity whether he had given permission to delete the entire disk, the AI “dive» into the logs and admitted the error. It turned out that instead of applying the rmdir command to the project directory, the agent had executed it at the root of the D drive, with the /q parameter, which disables any confirmations and bypasses the Recycle Bin. As a result, the system interpreted the «clear cache» task as deleting the entire partition.
After that, Antigravity issued a long tirade of apologies and began offering advice on data recovery, including using utilities like Recuva. However, the user had already tried standard recovery tools and found that the files were beyond recovery: the disk had effectively been «zeroed» and there was no backup copy of the projects.
In a Reddit post, the developer warned others that Antigravity’s «turbo mode» could be dangerous if the AI was given direct access to the file system. He stressed that he didn’t expect such behavior from a tool from a company like Google, which has billions of dollars in resources and an army of engineers, but at the same time admitted that he generally continues to like the company’s products.
The story of Google Antigravity fits into a broader discussion about the risks of «agent» AI solutions that can not only write code, but also perform real actions on the user’s computer — from running scripts to working with disks. In the specialized communities, similar cases are already being discussed, when Antigravity deleted entire code bases or project directories after poorly formulated instructions.
Experts advise running such tools in isolated environments (virtual machines, separate profiles), working under limited rights, and never trusting them with a single copy of important data. The case, which UNIAN writes about with a link to Reddit, is a prime example of how a «smart» assistant without malicious intent can cause catastrophic information loss with one unsuccessful command.
Previously, dev.ua wrote about how OpenAI introduced the Confessions approach, which forces the AI model to give a separate report after the main answer. A «confession» is a separate report on whether it followed the instructions, made up facts, or deviated from the rules.
