> And most of the potentially "bad" applications are also things you could do with a shell script.

I'll take this a step further - all of the current LLM attacks require a human to act on instruction. AI doesn't even provide a supposedly-secure platform like crypto does, it just issues completion through text. In that sense, it's not any different from those same impressionable humans visiting 4chan and reading something inflammatory. The medium is exactly the same, whatever you get out of AI could feasibly exist in the status quo before it.