We need more empirical research on biological AI model safety
Argues for more empirical research measuring the security-relevant capabilities of biological AI models, such as protein and genomic models, to inform biosecurity policy and funding decisions.