> OpenAI, for example, thought their sandboxes were good enough. As their AIs got more and more advanced, they kept proving them wrong - sandbox after sandbox.

Were their sandboxes in-process with the harness? Were they actually better than something like bubblewrap or even docker?