It was a painful search, but I finally found the video. I had an odd deja vu experience skimming these sandbox escape headlines. Almost a decade ago experts were already exploring the realistic expectations/implications of having/needing to sandbox ‘a malevolent AI’.
That discussion was awfully optimistic though. Sandboxing has inherit limitations, but increased friction usually incentivizes little to none. I still think ‘malevolent humans’ physically/quietly attached to a network are far more scary though. AIs are haphazardly rowdy, pushy, and loud.