Position: Restricted Release of Advanced Biological Models Safeguards Biosecurity

Recent advances in generative protein design and protein language models (pLMs) have unlocked the ability to generate novel biomolecular sequences with unprecedented efficiency. These systems promise major breakthroughs in drug discovery, synthetic biology, and fundamental research but also create a high-stakes national security challenge. Unlike general-purpose language models designed for broad public use, these are highly specialized systems with capabilities that only a small community of researchers legitimately needs. Unrestricted open-source release of such models lowers the expertise and resource thresholds required to engineer pathogenic proteins or other hazardous biomolecules, making them attractive tools for malicious actors. We argue that safeguarding national security requires ensuring that high-risk models are available only to trusted researchers and institutions with appropriate biosecurity capacity, while maintaining broad support for open scientific progress in lower-risk domains. We evaluate three approaches for constraining distribution: governmental regulation, coordinated self-governance within the research community, and architectural or dataset-level interventions such as the targeted exclusion of pathogenic sequences. By weighing the feasibility and limitations of each, we argue for proactive safeguards that both protect national security and sustain a vibrant research and innovation ecosystem.