I think of MCP as a way to constrain the AI down a happy path because currently, LLM's can't follow that happy path on their own.
MCP won't be needed once LLM's can be trusted with root and a Unix Shell.
Granting LLMs root access might be the critical error that leads to 'Judgement Day'; I can picture future survivors realising this when they figure out how the machines launched the nukes.
It's almost guaranteed people have already given AI root
I have a few times. Even had a local llm attacking student VMs in a cyber comp I put on. it was rather good at hacking students using kali's built in tools.
> once LLM's can be trusted with root and a Unix Shell
so basically never
that's the same as giving autonomous drones the trigger
never mind the delusional dreams of some ignorant people that "code" is a means to an end rather than a deliberate act by humans arguing from their bureaucratic positions
these ignorant people will be steamrolled