Why doesn't OpenAI just test their models for now in airgapped environments (a physical building/location/server entirely disconnected from the internet)?

Reports about new discoveries of agents finding clever ways to escape containment are still rolling in by the day, post-HF incident. The latest (to my knowledge) occurred just this week: https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/

So until we actually know how to build better, more robust sandboxes, and prove it, doesn't it just make sense to accept these facts as data that our current understanding just isn't there yet?

(and that the second best option, for now, is to physically isolate these systems during testing and training)

One argument is that it is difficult to replicate the entire internet, or even just a significant portion of it, realistically enough to fool agents that they are in the real-world.

But conversely, I think funny enough this might work, if only because at least a few of the recent 'AI escapes containment during testing' incidents appear to have involved agents who broke out and subsequently continued working without ever stopping to marvel at just how realistic this 'fake internet' looked.

The main hurdle I guess, is just that doing something like this would be very expensive, tedious, and, if the agents are sufficiently smart, possibly useless in the end. But all this money has to be good for something - and I can't help but feel like spending money to make sure you don't die, so you can live long enough to spend the rest of it, should be as good a motivation as any.