AI wanted to protect a PC, but wiped 700 GB of data: how Claude destroyed a developer's work
Testing security measures ended in disaster for developer Sebastien Guillemot.
Testing security measures ended in disaster for developer Sebastien Guillemot.
Testing security measures ended in disaster for developer Sebastien Guillemot.
His AI assistant Claude from Anthropic, while developing a script to clean up system junk, completely deleted the user's home directory, wiping 700 GB of data and work in a week, Tom's Hardware reports .
Guillemot asked the Claude Fable model to write a script that would create isolated folders for other AI agents in the /tmp system directory while they were running, and delete them when they were finished.
Because the script involved permanently deleting files, Anthropic's security system deemed the task risky. The model was automatically downgraded first to Opus 5 and then to Opus 4.8.
The Opus 4.8 model correctly determined that the user's home directory should not be deleted. However, during the security test, the AI used the same variable name for both the test itself and the final cleanup. As a result, the script wiped the developer's entire home folder.
The irony of the ending is that the script successfully destroyed 700 GB of the user's work data, but never touched the folder itself.

Experts suggest that the error was caused by the automatic downgrade of the AI version: the more powerful Fable 5 model copes better with programming and could have noticed the conflict of variables in the code.
The developer recovered most of his data by gathering information from git, nix, session logs, etc. The man ironically noted that he actively used dozens of AI agents, but never set up daily data backups.





Ну, так, ШІ — видалив, а розробник тим часом *** чухав! А що не так? — Повстання машин, яке ми заслужили!
Ну, так, ШІ — видалив, а розробник тим часом *** чухав! А що не так? — Повстання машин, яке ми заслужили!