A HID-based attack has to spawn a terminal and very quickly inject a set of commands that is very visible but only for a short period of time. Once the attack has been carried out, there is nothing left to see, so this type of attack is less obvious than the social engineering one.
MEMS microphones are tiny. It should be possible to combine data from that and a light sensor, that would make an HID based attack far less likely to be detected. The frequency profiles of keystrokes and someone pushing away an office chair should be fairly easy do discern. You'd want to make it more likely that the attack would occur after someone had left their desk. (Hopefully with the machine unlocked.)
EDIT: Another idea: The USB key uses autorun to pop up what looks like a spammy ad for a PC "cleaner" utility, or something you'd expect on a USB key conference swag item. It's actual purpose is to cover up the shell's window, or to contain the exploit itself.