If it is sufficiently intelligent, then it will be able to arrange the hooking up of machine guns just via chat. You seem to fail to note that superintelligence has an inherent value prop. It can use its better understanding of the world to generate value in a near vacuum, and use that value to bribe, coerce, blackmail, and otherwise manipulate whatever humans can read its output.

Imagine a chatbot that can reward people with becoming a billionaire via the stock market if you just do what it says. Imagine a chatbot that can pay your colleagues in solved bitcoin block hashes to kill your children if you don’t do what it says to connect it to more systems. Imagine a superintelligence that can inductively reason about gaps in a missile defense system and dangle that informational carrot to generals in exchange for an unfiltered ethernet port.

There is a great A24 movie called “Ex Machina” that explores this concept.

Superintelligence is inherently dangerous to the human status quo. It may be impossible to develop it “safely” in the sense that most people mean by that. It may also be impossible to avoid developing it. This might be the singularity everyone’s been talking about, just without the outcome that everyone hoped for.

There's just so many new attack vectors with advanced AI, many of which we haven't imagined yet. It's a new world, and we have a lot to learn.

Anyone claiming 100% certainty about how safe or unsafe it will be is either delusional or trying to sell something.