LLMs that work with audio data are apparently vulnerable to attacks using human inaudible commands because nobody who put together that LLM thought to impose a limit to the human audible audio frequency spectrum
Replies:
0
@campuscodi gonna be fun to include these in songs
@campuscodi "because nobody who put together that LLM thought-" evergreen statement right there
@campuscodi this restores my hope, as a vocalist I am soooo angry that bots are cloning voices. It seems that injecting prompts outside of human hearing could keep Ai from replicating what we hear at human levels.
@campuscodi oh my god, that's absolutely *fantastic*. It's time to PHREAK IT
@campuscodi it's not just LLMs and not just inaudible frequencies. Watch https://youtu.be/xMYm2d9bmEA
@campuscodi that sounds like them
@campuscodi considering that attacks on Alexa & Co using ultrasonic voice commands have been demonstrated years ago, this was pretty foreseeable.
Well, good that we are only putting these things in cars, nothing to see here...
