Testing Safety of Open Bio AI: Filtering Isn’t Enough
AI models for biology with downloadable weights promise faster drug discovery—but they also raise biosecurity concerns. A new study introduces a framework to test whether current “just filter the training data” safety practices actually work.
- What’s tested: Models’ understanding of viruses from three angles: reading genetic sequences, predicting mutation effects, and estimating how dangerous a strain might be.
- What they found: Filtering risky data often falls short—excluded knowledge can reappear quickly after fine-tuning and may even generalize more broadly.
- Why it matters: Signals linked to dual-use tasks may already reside in pretrained models and can be surfaced with simple probes.
Bottom line: Data filtering alone isn’t a safety net. The authors call for layered defenses and rigorous evaluations to guide responsible release of open-weight bio models.
Read more: http://arxiv.org/abs/2510.27629v2
Paper: http://arxiv.org/abs/2510.27629v2
Register: https://www.AiFeta.com
#AI #Biosecurity #OpenWeights #ResponsibleAI #Safety #Biotech #AIpolicy #MLSafety
