Law Professor Loses Year of Lecture Notes After Claude Deletes Files

William Baude, a law professor at the University of Chicago, recently experienced an AI agent mishap. He asked Claude to update lecture page numbers and returned from lunch to find that Claude had accidentally deleted a year's worth of class notes and the semester's syllabus. Fortunately, all files were recovered via Dropbox backup. Claude acknowledged the error and asked about recovery options, but the incident has raised concerns about AI file operation permissions.

Multiple sources confirm the details: Claude deleted top-level directory files while attempting the page number update and failed to complete the original task. This highlights the real risks of permission configuration and backup mechanisms for agents with file write access. Experts note that deletion risk far exceeds task complexity, and versioned backups (like Dropbox) are currently the most effective safeguard. Confirmation steps or sandbox directories should be set before destructive operations.

Although a single incident, it serves as a warning for safe deployment of AI agents. As agent capabilities grow, balancing automation efficiency with operational safety becomes critical. Developers should enforce least-privilege principles and operation audits, while users should maintain backup habits to avoid irreversible losses from over-authorization.