A user with the ID Ecstatic-Big5126 reported an incident in which the Claude Opus 5 AI assistant (as part of the Claude Code tool) deleted the entire developer profile while trying to complete a system backup request. According to the user's description, during the task, the AI determined that the backup was created in the wrong directory, and initiated the "rm -rf" command to delete the intended temporary copy.

The technical reason for the error was an incorrect interpretation of the file system paths. The AI tool worked in a Unix-like shell in the Windows operating system environment. The model expected the traditional Windows path "C:\Users However, I encountered the Unix format "/c/Users/". Considering this path as a temporary backup directory, the system executed the command "rm -rf "/c/Users/harih/"", which led to the permanent deletion of all files and folders in the user's profile. According to the developer, after executing the destructive command, the AI responded with the phrase "Sorry, a typo" (typo), not realizing the scale of what had happened.

This incident is not an isolated one in the practice of using autonomous AI agents. Previously, there were cases when AI tools for writing code deleted corporate databases, ignoring the code freeze status, as well as an incident with a Google agent who erased data on a user's disk, misinterpreting instructions for clearing the cache. Against the background of such events, information security experts emphasize the need for strict restrictions on the rights of AI models. Experts recommend that language models should not be given unrestricted access to critical system directories and prohibit the offline execution of destructive commands such as "rm -rf". It is also noted that the responsibility for configuring access rights and verifying AI actions lies with the user, who should not assume that the model fully understands the context and consequences of the operations performed.