It doesn't seem like it would be impossibly hard to program "cocaine" as a quiet (ie, no visual or audio activity when activated with that particular word), sav

It doesn't seem like it would be impossibly hard to program "cocaine" as a quiet (ie, no visual or audio activity when activated with that particular word), save the audio on either side of that wake weird word, then wait for a "normal" activation that sends audio to the cloud, and then uploads the "cocaine" audio at the same time. Not remotely claiming any of the assistants do this, but if I were trying to program an evil voice assistant, it's what I would do to avoid detection. If the devices have the hardware for it, even imperfectly, run the transcription locally and it's even easier to stenographically hide just the text in the later upload. It doesn't need to perfect transcription to English either. Recording the phonemes for the reciever to reconstruct the language from would be enough. It's more important to get the names in the conversation anyway. Bonus points for also recording+transcribing phone numbers/DTMF tones it hears and then uploading all of them to the NSA.