It is a little hard to see any actually happening, but I could imagine several scenarios:

• AI helping smaller scale actors build stronger tools of destruction

If it didn’t take nation-state level resources to build a nuke-level threat, easy to see how a few wackos could end us all. Could be bombs with less materials, bio-weapons, grey goo, whatever.

• Cyber security breaches enacting MAD

Presumably existing nuclear arsenals are well secured, but doesn’t seem beyond belief to imagine a hole that a persistent bot swarm could crack open.

• AI in hardware

We’re a ways away from sticking frontier models into small computers, but the push for robots is there. Once they have bodies in the real world, the limits are off for what harm they can cause. All the same unintended cyber attacks we’re getting from test environments now will then become physical harm, and with enough resources you could see it getting dire.

Typing those out though, does make me see how farfetch’d they kind of are. Which is good! I’d be all for it if the doomsday stuff turns out to be a big nothing.