Theory: probabilistic machines’ security is asymptotic: more parameters let it get closer to being secure/prompt-injection-resistant/whatever. It’ll never be perfect, but there’s some threshold beyond which it’s good enough.
To me this article reads as a celebration of how much better frontier models have gotten at defending against security flaws, rather than “open models bad”.
Eventually the tools we use everywhere will be “good enough to use and not worry”. This is foreign to software people, but only a Jedi deals in absolutes.