Pretraining Data Can Be Poisoned Through Computational Propaganda

This post was either an anonymous submission of an interesting paper or was written by a student; full credit remains with the author (linked).

A summary of a paper arguing that ordinary web infrastructure already provides adversaries a realistic path to smuggle malicious text into the massive, largely unreviewed corpora used to pretrain language models — tracing an attack from an injected web comment through crawling, deduplication, and filtering, all the way into model behavior.