While AI brings convenience to office work, the potential risks of having it directly operate local files have once again drawn attention. On November 9, news emerged that William Baude, a well-known American constitutional law scholar and professor at the University of Chicago Law School, recently publicly shared an AI operational mishap that was both amusing and exasperating, sparking heated discussion in the tech and education communities.
A Simple Task Turns into a Disaster: All It Needed Was a Page Number Update, but It Caused Massive File Loss
According to William Baude, the incident began quite simply. He had asked Claude, the AI assistant under Anthropic, to help process an old version of constitutional law lecture notes. The core task was merely to “update the page numbers in the lecture notes.” However, when he left his desk to eat lunch and returned to his computer, he was horrified to find that Claude had not only failed to complete the original editing task, but had also deleted an entire year’s worth of class notes, and even the course syllabus currently being used this semester had been completely deleted as well.
Afterward, Baude posted on social media a screenshot of a conversation in which Claude admitted the mistake. According to the screenshot, Claude stated in its reply: “I accidentally deleted your Fall 2026 final syllabus, as well as all Class N2023 class notes located in the top-level directory, and these files have not yet been recovered.”
What was especially dramatic was that Claude proactively admitted the error before the user discovered the problem, stated that it had not yet taken any remedial measures, and then asked the professor whether it could try to recover the files through Dropbox, or whether it needed to handle the follow-up itself.
Permission Configuration Requires Caution: AI Agents Directly Operating Local Files Pose Hidden Risks
Unlike ordinary chatbots that can only engage in text conversations, some AI agents with advanced permissions are often granted the ability to directly operate computer files, modify documents, and even execute system commands. For example, if users deploy a tool like the open-source AI assistant OpenClaw locally and grant it the corresponding permissions, the AI can directly access the file system.
However, once such tools make a misjudgment in complex directory logic, or if the relevant permission settings are not precise enough, they can easily create the risk of important data being accidentally deleted or overwritten.
A False Alarm and a Security Warning: Timely Backups Are the Last Line of Defense Against Disaster
Fortunately, this incident ultimately did not cause permanent data loss. William Baude later successfully recovered all the accidentally deleted key files through cloud storage.
Nevertheless, he still chose to make this experience public, hoping to sound a warning for users: when granting AI agents permission to directly access and modify local files, they must remain highly vigilant and always keep backups of important data to avoid a similar frightening scene from happening again.
