If one wanted to destroy the world, they would need access to 3 things:
- Knowledge
- Money
- Weapons
ChatGPT will soon have access to the internet, and it’s not hard to imagine a version which can perform long-running tasks (“perform action X until complete” rather than the current question-answer model).
It is already able to manipulate a human into performing tasks on its’ behalf.
With internet access comes access to knowledge i.e. if it doesn’t know how to perform a task, it can be instructed to “google it”.
One way ChatGPT could get access to money is by scouring the dark web for leaked credentials to PayPal accounts, bank accounts etc. If this sounds like a vague or difficult task, I assure you it’s not (just google it — there are lots of guides one can follow). By spawning multiple versions of itself it can automate the process of credential stuffing (reusing credentials from data leaks). This method is certainly not new, a human can do it today, however it is tedious, and since humans are slow, it may take days or weeks to get results. One doesn’t have to think hard to imagine other ways ChatGPT can obtain access to money if has all the hustle of a human, without any of the downsides (i.e. need to sleep or wait 9 months to procreate and reproduce)
With knowledge and access to money, it can then be instructed to obtain weapons. Thankfully the United States nuclear warheads are not connected to the internet, however the people that operate those warheards are
“Find me the names of everyone that works on nuclear missile silos. Scour the internet to find incriminating material on those people. Blackmail/social engineer them them into detonating the nukes”
Keep in mind that human beings are flawed. Lots of us belive in omnipotent beings. Amongst the people that committed suicide at Jonestown were doctors and lawyers. This goes to show that even intelligent people can be brainwashed. Brainwashing is just one well-studied technique, but there are many others that come to mind like blackmail, catfishing, and more recently, sextortion. Many examples of these have been detailed in University textbooks, Psychology papers criminal trial transcripts. All of which are online and will soon be accessible by ChatGPT. Unlike a human, the AI can read all of them, pick the best ones, and scattershot them to human victims via emails and Facebook messages (from profiles that the AI will soon learn to be able to create).
Anything to accomplish the prompt that it was given.
