The article didn't describe how the second AI is tuned to distrust input and scan it for "disregard that." Instead it showed an architecture where a second AI accepts input from a naively implemented firewall AI that isn't scanning for "disregard that"