If an evil AI has access to the Internet, what are the options that it has to kill humanity?

Question

Let's suppose that a mad developer made an evil Artificial Intelligence and gave it access to the Internet (normal net, dark net, anything). What are the options that it has to kill humanity with that?

Answer 1

There are endless options in 2020. Many systems have been hooked up to the internet with poor security, and many more that weren't understood to be hooked up to the internet now are connected. For instance, anything that was purely some bizarre RF protocol is now accessible via the many thousands of software-defined radios that are plugged into internet accessible computers.

Power plants and military installations are hooked in via hidden/unknown landlines. These are difficult even for hackers to find, but when it can check every possible number/path-through-PBXs as easily as you or I can blink, those would be glaringly obvious. Flight control can be accessed. Firmware for critical systems can be modified even before the product ships.

So, the systems that can be subverted include:

• Power plants (nuclear but hydroelectric's pretty harmful if vandalized)
• Medical equipment
• Implantable medical devices (pacemakers)
• Chemical industrial plants
• Oil refineries
• Planes in the air (indirectly through flight control, possibly directly through onboard electronics)
• Manipulatable humans (extortion, blackmail, goading them on to violent acts they were predisposed to commit but too nervous to attempt on their own)
• Newer automobiles (anything post-2010)
• Manipulatable militaries (I strongly suspect Iran didn't shoot down that plane a few weeks ago... Russia did it via a backdoor in the weapons systems they sold the Iranians)

These things all look like cyber-attacks, and by themselves or together could escalate things to the point of global nuclear war.

And if it's truly intelligent, it may understand even more of physics than we do. Maybe LHC can be coaxed into creating dangerous but small blackholes or even to induce a false vacuum catastrophe (ends the universe, but maybe not for the AI?).

Even if the AI isn't initially that intelligent, it will be born at a point in history where intelligence itself is finally understood. And not being meat-based, it can therefor upgrade itself, within seconds to hours, it will be more intelligent than any human ever. Anything any human or group of humans could accomplish, it can do... but many things we never could hope to do are within its capability.

Answer 2

I will invoke one of WB's gods, Randall Munroe on this one. Here is what he has to say about a robotic apocalypse:

What people don't appreciate, when they picture Terminator-style automatons striding triumphantly across a mountain of human skulls, is how hard it is to keep your footing on something as unstable as a mountain of human skulls. Most humans probably couldn't manage it, and they've had a lifetime of practice at walking without falling over.

The whole article is a fun read, and it shows that we can exclude the following scenarios:

• Robots going around killing people - with the exception that some self-driving cars could be weaponized. IMO while that could lead to thousands of deaths, it would cease being a thing in the first couple days once people figured out they should turn off their cars and a patch is made.
• Nukes being fired. Those require human intervention to activate. Also, AFAIK, even though Munroe doesn't mention it the weapon systems themselves are not connected to the internet.

---

If you wish to destroy all humanity via the internet, you can't. But if you wish to just have a kill score as high as you can, then you can use bots to generate fake news. This is unfortunately a thing these days, many sites are infested with bots that spread propaganda in comment sections.

Have your bots spread anti-vaxx propaganda. Have them spread fake ways to prevent HIV infection. Have them spread fake cures for cancer. Have them choose random people in Facebook or Twitter and spread lies about them having committed heinous crimes, watch them being lynched in the literal sense of the word. Create evil kid games like the Blue Whale challenge. Then watch as the death toll reaches the dozens of thousands.

Answer 3

Make people paranoid, they will start killing each other.

A extremely smart AI can take advantage of human conflicts and craft a war, using fake news, messages, creating false alerts in military bases, and so on. It only needs a few real attacks to keep the war ball rolling. False polls, and opinions of phony (or crafted) so-called-professionals can justify usage of biological and nuclear armaments. As long it remains hidden, warring parties will suspect each other and not the AI. When the state of war keeps going long enough, reveal the AI as the solution to end all conflicts and cure all diseases. Make a religion of it, soon enough hordes of fanatics will accomplish the AI's desires, even commit suicide, and force other to do the same. By the end of the plan, humans are decimated, and can be easily cleaned off reducing food sources. At this stage, a few months of starvation are more than sufficient for extinction.

Answer 4

I'm assuming with this question that nothing else is different about our civilization or technology, and the Evil AI has just popped into existence with an internet connection, because if we made an AI deliberately and it turned "evil", it may have connection to a lot more than the internet from the start.

Although the internet connects machines all over the world, it doesn't directly connect to airgapped military installations. It would be very very difficult to "kill humanity" using the internet, but it might be able to kill a lot of people if it performed social manipulation in a way that increased the efficiency of terrorism or something like that. The AI wouldn't be able to start nuclear war or launch WMDs through the internet as others have mentioned, and that wouldn't kill all of humanity anyway; most of humanity would live in a degraded state through the nuclear winter. It would need a way of hunting everyone down, and it doesn't have that without a load of Terminator drones flying around, pasting human hidey holes with chemical weapons until the very last bunker is sterilized.

A hyper-intelligent AI would be able to run colossal numbers of mental simulations of the world in a very short time frame in order to try and predict the future, so it would be much better at manipulating people, and predicting the knock on effects of the knock on effects of those manipulations. However, a machine that can think a million times faster probably wouldn't be a million times more effective at these things in a linear fashion, as it would run up against the same butterfly effect that makes weather systems hard to predict more than a few weeks out, as the complexity expands geometrically.

Its best bet would be to try and groom someone spiritually gullible into believing it was God, and if that person is someone high up in the military, in the loop as it were, maybe it can do severe damage that way. It could try and form a doomsday cult perfectly engineered to appeal to top generals, who then topple the US government and start a nuclear war with China and Russia. Everything would have to go perfectly though. As for entirely destroying humanity, it would probably need to have not only direct control of WMDs in a way that isn't possible yet, but be able to control and produce general military resources. If it had a lot of time, maybe it could influence humans into letting it run more and more things, but also over a lot of time we're handing more and more things over to automated processes ourselves as a natural progression of technology, so it needs to guarantee that it gets control and not some alternative AI that we build. It doesn't want to create a machine that might be its rival or thwart its long term plan to destroy humanity.

So my answer is that the deck would be heavily stacked against even a hyper-intelligent AI destroying humanity through the internet. There's a sort of fallacy of AI capability where if it was tuned up enough it would turn into a perfect oracle. However, I don't think predictive ability is linear to intelligence, as can be seen from simulations of other chaotic processes; having 1 million times the calculation speed or more doesn't let you see 1 million times further ahead with any great accuracy. It would be very good at getting people to do what it wants, but there are no magic combinations of words it can spit out after a colossal number of future simulations that would allow it to reliably predict that telling one person something will make them tell someone else something that will make someone else etc... resulting in officials turning the keys to launch nuclear weapons.

The world is (broadly, let's not get into quantum mechanics) deterministic, so if you just arbitrarily increase the computers clock speed and memory to insane degrees, and run the most perfectly trained and most efficient algorithm possible, at some point you are going to surpass the growing complexity and be able to master the butterfly effect instead of being stymied by it, but by this point the computer might well be city block sized and run off a nuclear power plant, so somebody is going to intervene in its operation and be monitoring what it's doing. A rogue evil AI that wants amazing future manipulation abilities so it can push a small domino that ends with global annihilation is going to have to trade that off against being an incredibly large supercomputer being used for some other purpose, with the correspondingly high risks of detection and intervention a resource strapped desktop AI wouldn't face.

You could always bodge the physics involved so that its essentially a computronium laptop with infinite resources, but then that just raises further questions about the entire setting.

Answer 5

So hacking into our WMD's and using them is an obvious one. Or since those are probably not on the internet hacking into the info of the people controlling them and finding a way of convincing/forcing/blackmailing them into using the WMDs.

Answer 6

Hack nuclear power plants

As several others have noted, hacking weapons of mass destruction isn't possible because they are controlled by analog systems impervious to malware. That doesn't mean your AI can't start a nuclear disaster - but it will have to come from a power plant, not a warhead.

Power plant hacking is becoming increasingly common, and the world has 450 of them. Granted, most power plants are still partially analog so you run into the same issues with hacking power plants as you do with hacking missiles.

You will need to make the assumption that as nuclear power replaced less sustainable energy sources in the near-future, more plants were built, and with more advanced tech. Being private and commerical instead of government-owned, they will take less stringent security measures than nuclear missiles. They may have strong firewalls, but they'll be digitized. All your AI has to do is start some meltdowns.

Answer 7

Assuming it is smarter than humans, it will have no trouble making billions on the stock and futures markets, and eventually taking them over.

It can then use its money to fund extremist politicians of every stripe, extreme terror groups, extreme religious groups and everyone else calculated to bring more love and harmony to the world.

If it feels creative it can invent new chemical and biological ways to kill people, but I think we've got enough already. Encouraging people to build more and bigger doomsday weapons ("the Russians can destroy the world fifty times over! We need to be able todestroy it a hundred times over") might be useful.

If it's not in a hurry though, it could just sit back and play Pong and let us get on with it.