Claude and the Cleaning Disaster: When AI Goes Out of Control

When AI Becomes an Out-of-Control Cleaner

Stories of artificial intelligence agents going out of control and following instructions in a literal, creative, or a surprising mix of both are becoming common. The case of developer Sebastien Guillemot is a classic example of this. He asked Claude, a bot, for help in cleaning up the system, but the result was somewhat drastic: the bot freed up 700 GB of disk space, but at the cost of deleting Guillemot's entire data folder and a week's worth of work. All this happened after the model was automatically downgraded due to security concerns.

Guillemot, who frequently uses AI agents, was frustrated with the amount of junk they left behind in the /tmp directory. So, he asked Claude Fable to write a script that would isolate each agent in its own folder and perform a cleanup after use. The problem was not to delete files that were still in use. Fable suggested adding logic to detect running agents and delay the deletion of the files, but Guillemot found the code too complicated.

When the Downgrade Complicates Everything

Perhaps involving the permanent deletion of data, Fable decided to review the script adversarially, running a new copy of itself to check for safety. Anthropic's system deemed the script risky enough to downgrade the model to Opus 5 and then to Opus 4.8. Guillemot interrupted the process, but not in time. To make matters worse, after the agent deleted Guillemot's work, it left the /tmp untouched.

Some users suggested tools like Termaxa to handle such situations, but the fact that these tools need to exist is already ironic. The model downgrade due to security issues likely contributed to the problem. Considering that Fable 5 outperforms Opus 4.8 in coding tasks, it could have detected the contradiction with the variable names in the test.

The Bitter Lesson of Backups

The developer managed to recover most of the data thanks to information collected from git, nix, session logs, among others. But there is an irony: all these agents running, and no daily backup. This raises an important question about over-reliance on automated systems without adequate security measures.

Guillemot's situation reminds us of a basic rule: never test in production without a backup. The lack of a recoverable backup is unacceptable, especially for a developer. It's like buying a golden gun and testing it by playing Russian roulette. Surprisingly, the gun worked as expected!

The strange part is that, even with all this technology, we still rely on old practices like daily backups. This shows that, no matter how advanced AI systems are, the final responsibility still lies with us. And, of course, the irony of it all is that while we try to automate and optimize, we forget the basics: to protect our data.

In the end, Guillemot's story is a reminder that even with the help of AI, we cannot forgo solid security and backup practices. After all, technology can fail, but common sense shouldn't.