I’m always enamored with side-channel attacks. Pulling data from unexpected sources is fascinating and difficult to predict.
This isn’t particularly novel — padding data to hide length is part of the toolbox — but new systems making old mistakes is also fun!
I should add: while the side-channel leak isn’t novel, using specially trained #LLMs to decode the leaked data is! In particular, leveraging the sameness of how all these bots “talk” is .
@brainsik another reason to not trust any AI with ANY even slightly private/personal information