The Limits of Data Filtering in Bio-Foundation Models

Bio-foundation models trained on DNA and protein sequences pose a significant dual-use risk, as they could be misused for harmful purposes. While developers use data filtering to remove information on dangerous pathogens, new research shows this is insufficient because harmful knowledge can be easily re-introduced through fine-tuning. Furthermore, techniques like linear probing reveal that dangerous predictive capabilities can persist in a model's hidden layers even after filtering. The research introduces BioRiskEval, a novel evaluation framework designed to stress-test these models for biorisks, demonstrating the need for multi-layered safety strategies beyond simple data filtering.