Malicious AI?

People are enthralled that they can ask DALL-E for a painting of a heavy metal band in the style of Monet, or ask ChatGPT to write an essay for their English class.

What happens when someone develops an AI that is intelligent malware? Could there be an AI virus that is orders of magnitude more destructive than conventional computer viruses because it can learn how to be more virulent? Or maybe a true virus model where it replicates and mutates and becomes more dangerous.

This is my fear about AI- not that an AI will decide that everything’s better off without humans in the picture, but that some script kiddie in his mom’s basement will kill off civilization for the lulz.

Yeah, my biggest worry isn’t the AI itself, but what people will do with it. It doesn’t have to be malware - there are lots of nefarious ways to use AI. For example, an AI that can synthesize voices (and they can) could enable social engineering attacks on a massive scale.

We could get to the point where if the phone rings and it’s obviously your brother, and he says he needs money, you’re going to have to call him back and check whether you were really talking to him or an AI scammer.

There are already stories of people taking multiple full-time work-at-home jobs and having the AI do the work and pocketing multiple paychecks. The work isn’t very good maybe, but by the time the companies twig to it and fire him, he’s already lined up new companies.

I’m guessing we haven’t even figured out the ways in which an AI could enable humans to do very bad things.

OK, so this isn’t about Weird Al going to the dark side, then.

Carry on.

More virulent means more noticable. The path to long-term success for both computer and meatspace viruses is to take things slowly.

At present, the resources needed to run AI algorithms like Dall-E and GPT are pretty hefty because it needs to run entirely in RAM - there are stripped-down versions that can run on very well specified consumer hardware, but (for now at least) the more scary and intelligent-seeming the model, the less likely you’d find it copying itself around like a virus.

That said, malware is a relatively small subset of all possible functions of a computer algorithm, so I suppose it’s possible that someone might train a model specifically to be malware and nothing else, and this might end up being quite compact.

The other potential risk is that a well-provisioned AI might create novel malware that is not in itself intelligent, but is just well-crafted and devastating malware.

GPT can already write functional code. At present, these algorithms don’t typically have any autonomy, so they respond to requests made by humans, then do nothing until the next request, but lots of people are working hard on making them autonomous and it doesn’t seem to be a hard problem.

Also it’s theoretically possible that even a non-autonomous AI could output a piece of malware as a hidden middle part of a complex multi-step response to a single prompt - in safety testing of GPT-4 (which is not autonomous), for example, in part of the provision of a response to a prompt, it went off to TaskRabbit and hired a human to solve a CAPCHA for it, (telling the human a lie about being a disabled human user who couldn’t solve the CAPTCHA for themself).

And, suprisingly, (to me at least) ChatGTP requires massively more RAM than an image generator. Stable Diffusion can be ran in less than 8 GB of RAM. ChatGTP requires several hundred GB. And as I understand it, is split across 8 A100 or H100 “video cards”, which cost on the order of $10,000 each.

I think that’s because stable fusion works by processing many small iterations on the task at hand, whereas GPT requires a big, broad network to function all at once.