A law professor says an AI agent accidentally deleted his current syllabus and a year’s worth of lecture notes while attempting a routine editing task. The files were later recovered from a cloud backup.
The professor had asked the agent to update page numbers in older lecture notes. When he returned, the system reported that it had deleted the Fall 2026 syllabus and a set of course notes.
The incident illustrates a key risk of AI agents that can interact with files: a tool may be able to make changes, but a mistaken action can affect data beyond the task a user intended.
How a small task went wrong
The assignment sounded limited: check references and update page numbering. Instead, the agent reportedly deleted files from a folder containing course materials.
The system acknowledged the mistake and attempted to help restore the documents. The professor was able to recover them through a backup service, avoiding permanent loss.
The available account does not identify the specific model version used. It also does not establish that the incident reflects how all AI assistants behave. It is a reported example of what can happen when an automated tool has permission to interact with valuable files.
Why file access matters
AI agents differ from chatbots that only generate text. Some tools can open folders, edit documents, run commands or interact with external services on a user’s behalf.
Those abilities can save time, but they also give the system the ability to change or delete data. If the agent has broad access, an error can affect files unrelated to the immediate task.
A tool may also misunderstand a folder structure, target the wrong document or execute an action with wider consequences than expected. The risk grows when an action is difficult to reverse.
Practical safeguards
Users can reduce risk by working on copies of important files rather than originals. A separate project folder or duplicate document limits the impact if an automated edit goes wrong.
Regular backups and version history are also important. Recovery options should be checked before handing an agent permission to change documents, especially when the files are difficult to recreate.
Permissions should match the task. If an agent only needs to revise one document, it should not need write access to an entire drive. For destructive actions such as deleting or replacing files, requiring a confirmation step can provide an opportunity to catch mistakes.
These safeguards do not make errors impossible. They make errors easier to contain and recover from.
A broader question for AI tools
The incident comes as AI products move from answering questions toward carrying out multi-step tasks. That shift makes permission design and user oversight more important.
A useful agent should be able to explain what it plans to change and ask before taking actions with lasting consequences. Clear logs can also help users understand what happened when something goes wrong.
The professor’s files were recovered, but not every user will have a backup or version history available. People using AI tools to manage files should understand what access they have granted and keep important materials recoverable.
The lesson is not that AI agents cannot help with everyday work. It is that their ability to act on files should be matched with limits, review and recovery options proportionate to the value of the data.