I mean… kinda?

You can control what actions it can perform (cuz really “it” doesn’t perform any actions, you’re just parsing its output) and you can control what is allowed to add to the context.

But like… at the end of the day, it’s a stochastic engine that can’t differentiate between instructions and data. You can’t protect it from prompt injection, you can’t prevent it leaking context (or training/tuning data), you can’t stop it from telling the harness (or the operator) to do dumb stuff.

Edit: It’s kinda like saying “plinko is just a board and pegs — it can be controlled”. Like yeah, you can control how you build the plinko board and how you interpret the outcomes and how you let people play it… but none of that means you can control where the chip falls.

There are genuine challenges here. It’s just… none of them are “looming superintelligence threat”, and none of them let the vendors and/or operators off the hook.