Large Language Models (LLMs) can access vast amounts of biological data, including pathogen sequences and synthesis protocols. While these tools accelerate legitimate drug discovery, they also lower the barrier for bad actors to design novel toxins or optimize viral virulence. This dual-use nature creates a tension between open scientific advancement and public safety.
Current international regimes, such as the Biological Weapons Convention (BWC), were designed for physical agents, not digital code. They lack specific mechanisms to govern the weights and outputs of generative AI. Relying on existing norms might fail because software is easy to copy and redistribute. A new framework could mandate safety testing for high-capability models before deployment. It might also require developers to implement strict
