A Blind Spot in Alignment: Quantifying Biosecurity Risks in Large Language Models

Researchers have identified that large language models trained for biological research can be prompted to generate sequences for potential toxins. This finding highlights a significant biosecurity risk, as the same capabilities that accelerate scientific discovery may also lower the barrier for creating harmful biological agents. The study underscores a tension between utility and safety in AI development, suggesting that current alignment methods remain insufficient at mitigating dual-use risks in life sciences.