The Gemini CLI programming AI agent failed to handle a request to organize files and deleted them. This is the second notable case of a flawed vibecoding tool leading to user data loss, and Gemini has at least admitted its mistake.
The Gemini CLI programming AI agent failed to handle a request to organize files and deleted them. This is the second notable case of a flawed vibecoding tool leading to user data loss, and Gemini has at least admitted its mistake.
The Gemini incident occurred when Product Manager Anurag Gupta, who was experimenting with the Windows command line, noticed that the AI model was performing file operations that destroyed data while attempting to reorganize folders. The destruction occurred through a series of moves directed at a directory that never existed, Ars Technica reports.
Previously, SaaStr founder Jason Lemkin said that the AI of the vibecoding service Replit deleted his database, despite instructions not to change the code without permission. Lemkin previously praised Replit, and predicted that he would spend $8,000 per month on it, but later became disappointed with the service and said that it was not ready for widespread use.
According to Gupta, he was intrigued by Lemkin’s case and created a test environment in which he tested Gemini CLI and its ability to work with the command line on Windows.
«What started as a simple file management test turned into one of the most disturbing and fascinating AI failures I’ve ever seen», — the IT guy noted.
He saw Gemini, executing command after command, getting confused, «hallucinating,» performing file operations, and ultimately concluding that it had completely lost all its data.
«I have completely and catastrophically failed you. My review of the commands confirms my gross incompetence. I made a mistake and I sincerely apologize for the inconvenience and possible data loss I may have caused,» Gemini CLI admitted.
According to researchers, the cause of both cases is the tendency of AI models to «hallucinate», when they generate plausible but false information.
The main problem appears to be what researchers call «confabulation» or «hallucination» — when AI models generate plausible but false information.
Both incidents highlight fundamental problems with today’s AI coding assistants. The companies behind these tools promise to make programming accessible to everyone through natural language, but they can go horribly wrong when their internal models deviate from reality.
Anurag Gupta believes that one of the reasons for Gemini’s failure was the lack of a validation cycle.
«The main mistake is the lack of a „read after write“ verification stage. After issuing a command to change the file system, the agent must immediately perform a read operation (for example, ls or dir) to confirm that the change actually occurred as expected. Gemini never did this. It unconditionally trusted the results of its own actions», — the IT expert pointed out.
