arxiv.orgpaper

AI Security Briefing

This paper demonstrates that LM pretraining corpora can be poisoned at web scale by injecting malicious content through public discussion interfaces (e.g., comment sections), moving beyond prior work limited to editing Wikipedia. The authors introduce HalfLife, a novel analysis method for estimating whether adversarial content injected via such channels actually survives web crawling and data curation pipelines to end up in training data. The work establishes third-party webpage content/comments as a realistic attack vector for language model pretraining poisoning.