Downvoting out of fear.
LLMs are about as safe as the internet in 1993. Tampering with the training data to inject bias is a massive attack surface that is being ignored. It makes sense: trying to address corruption of models this way would…
Downvoting out of fear.
LLMs are about as safe as the internet in 1993. Tampering with the training data to inject bias is a massive attack surface that is being ignored. It makes sense: trying to address corruption of models this way would…