I think it's good that someone is making this point, anyway. For sandboxing a super-capable offensive-security AI, you would think that cloning PyPI and running

I think it's good that someone is making this point, anyway. For sandboxing a super-capable offensive-security AI, you would think that cloning PyPI and running it as an offline service ought to be table stakes, but apparently that's not how OpenAI saw it, for instance.