Just take the whole OpenAI / Huggingface thing for a starter. Remember, I only said "cause great harm" not "eliminate all humanity". Within that rubric, I think that incident, and other recent similar incidents, already show the plausibility of AI systems doing unexpected and harmful things. Being automated, I think it's implicit that they can easily scale, so I have no problem making the leap to "cause great harm."

Beyond that, we already know AI systems have caused harm in various contexts where they have been used, like facial recognition getting people arrested without justification, loan applications being denied wrongly, jail sentencing being mishandled, and so on. All of these things have actually happened in the real world already. So it doesn't take much imagination to extrapolate a little and see the potential for greater degrees of harm. And consider the standard we're comparing against here: "probability greater than zero". That's not a very high bar, so I feel pretty well justified in holding this position.

What does "based on reality" mean to you?

You have poor epistemics if you are still assigning exactly 0% probability given developments of late.