It's nice to see that the "flash crash" as a key risk from autonomous AI agents get popular unlike when I used to talk about it a few years ago, eg it came up in Odd Lots just now; it's been my key unforeseen risk vector beyond normal misuse concerns