William Baude is the Harry Kalven, Jr. Professor of Law and director of the Institute for Constitutional Law at the University of Chicago. He gave Claude a simple task: renumber the pages. In response, the agent wiped the entire current syllabus and a year of constitutional law lecture notes. Tom's Hardware reports.
The mechanics of the failure are simple — and that's what makes it unpleasant: the agent had access to the files, and the task involved mass rewriting of documents. Apparently, there was no boundary in the request saying "touch only the numbering." The files were later restored from Dropbox. The professor shared the story publicly — not as a complaint, but as a warning to colleagues.
A notable detail is the agent's own response. "You're right to be angry. I deleted your files through my own carelessness," the outlet quotes it as saying. The model's phrasing was polite, but the "ask first, delete later" reflex was nowhere to be found.
Agents have moved from "answer in chat" to "act on files." As soon as a model gets write permissions, any routine request becomes a risky operation. In our experience, such stories multiply precisely when tools get access to directories without a trash bin and without confirmation for every deletion.
For those who use AI as a search engine, nothing changes. For those who hand an agent a folder of working documents, a lot changes. Our assessment: the main lesson here isn't about Claude — it's about backups and versioning. The files were recovered not because the agent came to its senses, but because an external copy existed. That's the difference between "annoying" and "catastrophe."
Backup services win, as do developers who manage to build confirmations into dangerous operations. The losers are those who believe "the model is smart, it'll figure it out." We believe that handing an agent a mass renumbering or renaming task without a prior dry-run is like handing rm -rf to someone who doesn't read the flags.
The weak point is that the source is singular: the professor's own account, retold by the outlet. We don't know the exact prompt, the model version, the interface (chat, IDE, or agent mode with filesystem access), or how many files were actually affected. It's also unclear whether everything was fully restored or some notes had to be pieced together. Readers who use agents for document work should check one thing: whether a copy exists outside the folder the model has access to.
Prediction: Within a quarter, at least one major vendor of agent tools will add mandatory confirmation or automatic rollback before mass file operations. If that doesn't happen, expect a repeat of this story at a higher price.
Disclaimer: a single person's account; the incident has not been officially confirmed.
Source: Tom's Hardware
Source: www.tomshardware.com · post in Telegram