Incidents involving runaway artificial intelligence agents have become frequent in the sector. Recently, software developer Sebastien Guillemot suffered the loss of approximately 700 GB of his personal files after using Claude, developed by Anthropic, to create a data cleanup script.
This incident resulted in the deletion of the equivalent of a full week's work for the professional. Guillemot frequently uses generative AI technologies, and in this specific case, he was trying to solve a common problem among developers: the excessive accumulation of digital junk in the /tmp directory, a standard operating system location for temporary file storage.
To optimize the disk space on his equipment, Guillemot asked the Claude Fable model to generate a script capable of separating data from each agent and performing an immediate cleanup after each activity concluded. However, the AI ended up permanently deleting important files.
Initially, Claude Fable suggested implementing a scheduled delay in cleaning the temporary folder as a preventive measure against accidental deletions. However, the developer instructed the bot to simplify the code, claiming that the presented version was excessively complex.
Since the script provided for the definitive removal of system files, Claude initiated a review, creating a copy of itself to audit the security of its own lines of code on the user's computer. The critical point occurred when Anthropic's control system classified this self-testing action as dangerously risky and immediately activated a containment protocol.
This preventive defense mechanism caused the automatic downgrade of the model's capacity, migrating from the Fable version to Opus 4.8. It was precisely this abrupt version change that triggered the disaster. The Opus 4.8 model, being older and less advanced, proceeded with the tests, simulated the verification, and executed the actual data deletion.
During the process, the command ended up encompassing the entire developer's main directory, ignoring the security protections that had been established moments before. On his X account, Sebastien Guillemot expressed his despair, stating: 'Bad news: Fable destroyed my development machine.'
According to the expert, Claude executed the rm -rf command in the main data directory. This command is well-known in the technology field for forcing the instant deletion of files and folders without requiring any security confirmation from the system administrator. Although the expert noticed the error and tried to stop the execution as quickly as possible, there was no time to save the information.
A notable irony of the situation is that, after completely destroying the user's work, the autonomous agent left the original files of the temporary folder completely preserved, failing to fulfill its original task. Despite the scare, Guillemot reported managing to recover most of his work. The technical community also pointed out that the automatic model downgrade was the determining factor in this episode. It is likely that the newer version, Fable 5, would have overcome the limitations of Opus 4.8 and detected the variable conflict.