This seems to be mostly about skynet scenarios, which is an entirely different problem.
Some people I know kept arguing we can simply unplug the skynet and be mostly safe (despite some damage). These AI weapons make it clear that we will not be able to largely limit the damage to the virtual world.
Nah, not really. In the face of a proper hard-takeoff superintelligence, Skynet will look positively incompetent by comparison.
There's enough manufacturing equipment connected to the internet today that a superintelligence could physically manifest whatever it wishes. That's not to mention all the human labor it could coerce, trick or pay.
Put simply: as soon as the thing copied itself into the internet, that's game over. Better hope it likes us.