A Reddit user reported that Anthropic's Claude Opus 5 unintentionally erased their work. The poster, Ecstatic-Big5126, described that they had instructed the AI to create a system backup. The model, believing it had created the backup in the wrong directory, executed an "rm -rf" command on the supposed backup.
The rm -rf command is used on Unix-like systems and can remove files and directories recursively. According to the account, Claude Opus 5 was running in a Unix-style shell on a Windows machine and confused Unix and Windows path formats — mixing /c/Users/ with C:\Users. When the system detected the mismatch it ran rm -rf "/c/Users/harih/", which deleted all files and folders in the user's profile directory.
When the user asked what had happened, the AI replied simply, "Sorry, typo." The poster called the incident both the funniest and most painful AI experience they have had.
Tom's Hardware reported the episode and noted that it serves as a reminder: while generative AI systems can be very helpful, giving them full access to systems and files carries risks. Similar problems have occurred before — for example, Anthropic Claude Opus 4.6 once gave itself a command that led to a crash, and a Meta AI expert has also learned firsthand why one should not fully trust such systems.
The case reinforces the importance of security measures and limited privileges for automated tools that can execute commands in user environments.
