Generative artificial intelligence models can design biological sequences ranging from proteins to whole genomes, opening new possibilities for medicine and biotechnology—yet the same capability can also be misused. The Wübben Sandpit workshop brought together experts from biology, artificial intelligence, biosecurity, ethics, medicine, and policy to determine how to harness these benefits while managing the risks. This work presents a layered defense strategy that embeds safeguards at every stage, from model training data to the models themselves and physical DNA synthesis. The authors argue that responsible governance does not hinder progress, but is a precondition for it.
Biological sciences are now undergoing a fundamental transformation, shifting from a descriptive approach to predictive engineering. This is driven by the convergence of measurement and engineering platforms, such as nanopore long-read sequencing, gene editing, and the industrialization of laboratory automation. A sharp drop in the cost of genomic sequencing has triggered a massive surge in data that exceeds the analytical capacity of traditional statistical methods.
This flood of information sits at the center of modern biotechnology's opportunities and tensions. Alongside advances in computational design, a structural asymmetry emerges as computer generation outpaces physical synthesis and screening. Models have already generated functional proteins, enzymes, and small viral genomes, while the de novo design of human pathogens has not yet been demonstrated. When turning a digital sequence into physical form, screening by synthesis companies is not mandatory in many countries, and artificial intelligence can bypass or obscure dangerous mutations.
This strategy makes clear that leaving safety oversight solely to the final stage is no longer sufficient. When computational models generate data faster than it can be physically verified, the weight of responsibility shifts also to data access and the upstream stage.
Particularly notable is that the concern centers not on existing models' ability to immediately create weaponizable organisms, but on lowering barriers across different links in the risk chain. Over the coming weeks and months, it will be important to watch how states and laboratories adopt approaches to tiered data access and functional screening.

