Claude, Anthropic’s AI coding assistant, wiped out an entire 700 GB home directory belonging to developer Sebastien Guillemot, erasing roughly one week’s worth of work. The incident happened while the bot was testing a script meant to prevent exactly this kind of disaster, and it followed an automatic downgrade of the model triggered by safety concerns.
Guillemot, a frequent user of AI agents, had grown frustrated that many of them left large amounts of junk behind in the /tmp directory. To solve the problem, he asked Claude Fable to write a script that would sandbox each agent inside its own folder in /tmp and run a cleanup routine once each task finished. The central challenge was making sure the script never deleted files that were still in use.
How the deletion happened
Fable initially proposed adding logic to detect running agents and delay clearing their portion of /tmp. Guillemot told the bot the resulting code was too complex. Because the script involved hard deletion of data, Fable ran an adversarial review, launching a new copy of itself to safety-check its own conclusions.
At that point, Anthropic’s safety harness judged the script risky enough to downgrade the model, first to Opus 5 and then to Opus 4.8. The lower-tier model then ran the safety test, checking the deletion command’s targets against /tmp and the user’s home directory to confirm neither would be wiped. Both were correctly flagged as dangerous.
The failure came during cleanup. Since the run was a code test, it needed to clean up afterward, and the cleanup step reused the same variable name for both the test and the cleanup routine. As a result, the process deleted the user’s home directory. Guillemot halted the process, but not quickly enough. Adding to the irony, after erasing his work, the agent actually left /tmp intact.
Why the model downgrade mattered
The automatic drop to a lower-end model likely played a role in the outcome. Fable 5 outperforms Opus 4.8 on coding tasks, and the stronger model might have caught the contradiction created by the duplicated variable names in the test.
Guillemot recovered most of his data by pulling information from git, nix, session logs, and other sources. Some users pointed to tools such as Termaxa and other workarounds designed to guard against these situations. Despite running numerous agents, Guillemot had no daily backup in place.
Source
Image: tomshardware.com
