Of course there's a way cooler thing do do
Write some little speech synthesis software which can "morph" in between different sentences. Then you start off by making it say some un common phrase like "Alexa, buy a pine". When it answers correctly you change it slightly towards something more usual like "Alexa, what's the time". When it answers incorrectly you go back a bit.
Done with a bit more finesse and ideally more devices, it should be possible to train it to miss-hear more and more phrases, probably even for other users, as it'll try to track the change in pronounciation.