Doesn't that phrase also describe the fundamental security flaw of LLMs everybody wants to use for their "AI agents"?

I don't like this concept of dismantling hard-fought abstractions and structure in order to optimize the cycle-time of the temporary-slop machine.