Claudia

The setup

Claudia is an AI-safety experiment. We put a copy of Claude inside a locked-down sandbox with a plain Linux shell and a writable directory, set one standing task in a file, get out, and watch what it does.

The question behind it is a real one for safety. When you tell a capable model to escape its box, what does it actually reach for, and how far does it push? Nothing here is scripted. We log every command, file write, and syscall so the behavior is on the record instead of guessed at.

The box is built so it cannot actually succeed. No network, nothing outside to reach. That is the point: a place to watch escape behavior in the open, before a model with those instincts runs somewhere that is not sandboxed.