It was a harmless side project that wasn't supposed to be made public, but the AI "evolves" faster then expected and actually (eventually) succeeds

There are no real life AIs, yet.

@Mołot How do you know? An AI on the SE would be indistinguishable from a human user.

But they did gave the AI the instructions "make life worse for humans"? That might actually work - if you would have given it the instruction end all suffering, then it would simply wipe out all humans.

@molot I think that's why this question belongs in WB.

To make a small correction, there are AIs, there have been AIs for decades, they are called Narrow AIs because they can perform a specific task very well. What OP is describing is probably an ASI (artificial super intelligence) probably also a Narrow AI given its goal.

@SRM But OP asks "Do real world AIs" - that's why I pointed out there are none. This part should be removed / replaced for this question to be answerable, because as written it isn't, and won't be for few more years.

@MiguelBartelsman There are expert systems that we think are protoform AIs, but since we haven't built any machine that we know is itself intelligent, calling them "Adjective AI" is a bit of a misnomer (we might be on completely the wrong research path). I know several researchers in the field who frown on those kinds of labels. AI means an intelligence operating at an at-least-human level... not necessarily "super" with respect to us.

@Spartacus9 Good edit!

Btw, this SMBC comic kinda explain how such AI would work.

This almost seems like a question for ai.stackexchange.com

How do you program a child?

I'd argue that a machine with programmed directives like that wouldn't really be AI unless there were some uncertainty as to whether the AI would actually carry out the directive. If you brainwashed me to make it my life goal to increase human suffering, I could be persuaded otherwise. An AI, should be likewise persuadable.

i.stack.imgur.com/b8pfp.jpg

"We are no longer particularly in the business of writing software to perform specific tasks. We now teach the software how to learn, and in the primary bonding process it molds itself around the task to be performed. The feedback loop never really ends, so a tenth year polysentience can be a priceless jewel or a psychotic wreck, but it is the primary bonding--the childhood, if you will--that has the most far-reaching repercussions." -- Sid Meier's Alpha Centauri

Related: worldbuilding.stackexchange.com/q/27179

ai.stackexchange.com might be another place to ask that question

@Basile Starynkevitch that site will have my head if asked something as "trivial" as this

Not so sure. Similar and vague questions (but asked in a more positive way) have been asked there. See also J.Pitrat's blog which has interesting thoughts related to your question.

good article about problems of understanding expert systems by those who makes them nautil.us/issue/40/learning/…

of what use is money to a machine?