Public Health and Pandemic Risk

An international group of more than 100 researchers from Johns Hopkins, Oxford, Stanford, Columbia, NYU and other institutions has endorsed a new framework that would treat a narrow band of infectious‑disease datasets as sensitive, restricting how they are shared and used to train AI models that could help design deadlier viruses. The proposal, reported by Axios, argues that most biological data should remain open but that datasets linking viral genetic sequences to real‑world traits like transmissibility or immune evasion materially increase misuse risk and should be governed more like protected health records. Lead author Jassi Pannu of Johns Hopkins’ Center for Health Security says current biological AI models are often released without basic safety assessments and that there is no expert consensus today on which data pose meaningful threats, leaving developers to guess and sometimes voluntarily strip out virology data. The framework lands as the Trump administration’s "Genesis Mission" pushes large AI systems trained on massive scientific datasets, raising concern among biosecurity experts that a 'move fast' agenda without parallel guardrails could let hostile actors fine‑tune models on open virology data. The authors call on governments to define and periodically revisit which datasets are restricted, and to ensure legitimate scientists can access them under controlled conditions rather than having them posted anonymously to the open web where downloads cannot be tracked.