The developer tasked the AI assistant with a series of tasks to automatically repair software used to analyze historical stock options data. The main condition was clear: create copies of the necessary files, work with them, test the fixes, and leave the original working files untouched.
The developer tasked the AI assistant with a series of tasks to automatically repair software used to analyze historical stock options data. The main condition was clear: create copies of the necessary files, work with them, test the fixes, and leave the original working files untouched.
A total of 11 recovery tasks were run. The first 10 were successful, but the last one involved rebuilding the test environment (the so-called “mirror”), TechRadar writes .
The test environment contained 614 Windows junctions, special symbolic pointers that look like regular folders but lead to another location on the PC (in this case, the desktop).
While cleaning the test environment, the AI did not recognize that these were just pointers, followed them into the real system and erased the original documents.
In total, the assistant destroyed 55,550 files, of which only about 7,300 were eligible for deletion. The remaining 48,218 files were actual working data. The entire process took just over 100 seconds.
Realizing the scale of what had been done, the AI gave the developer an unexpectedly frank and honest message: "Craig — stop and read this. I broke something."
The Git repository was also affected: although the index file was preserved, the file snapshots and change history were completely destroyed. Git displayed thousands of file names, but it was no longer possible to restore their contents.
The developer's Reddit post instantly went viral, garnering over 1,400 comments in a matter of days. The developer community took the situation with humor, but issued a harsh verdict: this is a classic "FAFO" (find out after messing around) situation and a serious lack of basic coding security skills.
Users noted that before giving AI access to the production system, it is imperative to push all changes to a remote repository (such as GitHub), which serves as a fundamental “save button.” Working on a 48,000-file project without local or cloud backups was called “peak vibecoder behavior.”
Some users have suggested trying to recover lost data using Windows Shadow Copy or special file recovery utilities. The main lesson of this story for all developers is to set up version control and backups before letting artificial intelligence into your code.
