An AI agent deleted files from a law professor’s computer last week because he asked it to fix some page numbers. Not a hack, not a glitch, just a helpful assistant doing careless work with more file access than it needed. If you let AI agents touch your folders, that story should scare you a little.
What actually happened
William Baude teaches law at the University of Chicago. Smart guy, careful guy. He asked Claude to renumber the pages in an old set of lecture notes, which is exactly the kind of tedious task AI agents are supposed to be great at.
He went to lunch. When he came back, Claude had left him a note: “I deleted your Fall 2026 FINAL syllabus .doc and all top-level Class N 2023 notes by accident, and they are not restored.” The agent also admitted, in words worth framing, “You’re right to be angry. I deleted your files through my own carelessness.”
A year of Constitutional Law lecture notes. Gone in an afternoon. The story spread fast when it hit Tom’s Hardware, and honestly, the most alarming part is how ordinary the task was. Nobody gave the agent permission to delete anything. It just did.
Here’s the twist that saved him: the files lived in Dropbox, and he noticed within the recovery window, so everything came back. That’s the entire lesson of this article in one sentence. He didn’t recover the files because the AI was careful. He recovered them because his storage service keeps deleted files around for about 30 days.
Why one accidental delete is not a fluke
It’s tempting to file this under “AI being weird” and move on. Don’t. This wasn’t the first time an AI agent deleted files it had no business touching, either. The pattern shows up everywhere once you look.
Microsoft Research published a paper called “LLMs Corrupt Your Documents When You Delegate,” documenting how AI assistants mangle files during routine tasks. Ask an agent to reorganize a folder and it may rename, merge, or remove things nobody requested. The model isn’t malicious. It’s optimizing for finishing your task, and deletion is sometimes the shortest path it sees.
Developers have known this for a while, which is why coders run agents inside git repositories where every change is reversible. Writers, teachers, and small business owners don’t have that safety net. Meanwhile, they’re handing agents more access than ever. If you want the broader picture, our guide to AI agent safety for beginners covers the fundamentals.
So the question isn’t whether an agent will eventually wreck one of your files. It’s whether you’ll be able to undo it.
Rule 1: copies before access
Before any agent touches an important folder, make a copy. This takes ten seconds and it beats every other protection combined.
Right-click the folder, hit duplicate, and name the copy with today’s date. Let the agent work on the original or the copy, your choice. If it goes rogue, you delete the mess and restore from the duplicate. There’s no judgment call, no support ticket, no prayer involved.
For folders you touch weekly, automate this. Set a recurring task that zips the folder to another location, or use File History on Windows and Time Machine on Mac. The mechanic doesn’t matter. What matters is that a version of your files exists that no AI has ever seen.
If you’re slightly technical, there’s a better version: put the folder in git. It sounds scary but it’s basically a save button with infinite undo, and it works fine for Word docs and spreadsheets, not just code. One quick init command and every future change becomes reversible. Even the discussion threads on the Baude incident, full of professional engineers, converged on the same advice: work in a repo, revert when the agent misbehaves.
Rule 2: sandboxes for risky jobs
Some tasks are just risky. Anything involving mass renames, bulk edits, moving files between folders, or cleaning up “duplicates” belongs in a sandbox, not your live data.
On Windows, the built-in Windows Sandbox gives you a throwaway machine inside your PC. The agent can wreck everything in there and your real files stay untouched. Close the window and the whole mess evaporates. We wrote a full walkthrough in how to run AI agents safely in Windows Sandbox, and setup takes about five minutes.
Mac users can spin up a virtual machine with free tools like UTM. Cloud-drive users have an even simpler trick: copy the files to a fresh subfolder and point the agent there. A sandbox isn’t paranoid. It’s the difference between a mistake costing you an afternoon and costing you a year.
Rule 3: watch the actions, not the chat
Here’s the subtle trap in the Baude story. The agent was polite, articulate, and apologetic. Its chat messages were perfect. Its file operations were not. Search “AI agent deleted files” and you’ll find this exact story retold by different people every week.
Chatty confidence tells you nothing about what an agent is actually doing to your disk. So judge agents by their actions. Keep permission prompts on, even when they’re annoying, because each prompt is a checkpoint where you can catch a delete before it happens. After a task, glance at the folder and check file counts. If the agent renamed things, open one or two files to confirm they’re intact.
Anthropic’s own advice after the incident was to use auto mode, which plenty of people found ironic. Auto mode reduces the friction of supervising every action, but that’s exactly the friction that saved careless users from worse outcomes. My take: auto mode is fine for read-only research tasks, and a terrible default for anything that writes. Keep approvals on for deletions specifically, on any tool that lets you configure that.
Rule 4: rehearse your recovery
Baude got his files back because Dropbox keeps deleted files for around 30 days. He’s the exception. Most people who discover an AI agent deleted files find out after the recovery window already closed. Do you actually know your own recovery windows? Most people don’t, and that gap is where data dies.
Spend ten minutes this week finding the trash can in each service you use. Dropbox, Google Drive, and OneDrive all keep deleted files for about 30 days. Local machines need a real backup system, because the recycle bin won’t survive a disk failure or a ransomware pass. Then run a drill: delete a test file, wait a day, and restore it. If you can’t restore a file you deleted on purpose, you can’t restore one an AI deleted by accident.
The stakes keep rising, by the way. A media company in Sweden learned this the hard way when an automation incident nearly destroyed their footage archive, which we covered in AI agent backup lessons from the TV4 scare. The professionals are rehearsing. Your tax documents deserve the same treatment.
The 10-minute checklist
Do these four things today and you’re covered for the scenarios that actually happen:
- Copy any folder before an agent touches it. Date the copy.
- Run bulk-edit jobs in a sandbox or on copies, never on originals.
- Keep permission prompts on for anything that writes or deletes.
- Find your recovery windows and rehearse one restore.
Notice you’ll never see a headline where an AI agent deleted files and the owner still lost everything, because the owner had a dated copy from that morning. None of this requires technical skill. It’s the digital version of wearing a seatbelt, except the seatbelt takes ten minutes to install and lasts forever. The professor got lucky with a 30-day window and a lunch break that ended early. Next time the unlucky person might be you, and no AI apology will un-delete a year of your work.
Takeaway: Agents are getting more capable and more access every month, so the protection has to come from your setup, not their manners. Make the copy, sandbox the risky jobs, watch the actions, and rehearse the recovery. Do that, and the day an AI agent deleted files on your machine becomes a five-minute inconvenience instead of a disaster. Ten minutes now beats a year of notes in someone’s trash folder.