The Python it writes runs in a sandbox environment with its own containerized filesyatem. That's been obvious to me for a while as I have seen it generate temporary files when uploading things like images to scan.

This does not imply that any large language model executes the code it writes... I get it, they don't. I build models myself and have a fairly good understanding of how they work. But ChatGPT-4 does provide the ability to generate and run code in its own sandbox environment, outside of the model itself.