Attention Track: Without Safeguards, AI-Biology Integration Risks Accelerating Future Pandemics

Dianzhuo Wang, Marian Huot, Zechen Zhang, Kaiyi Jiang, Eugene Shakhnovich, Kevin M. Esvelt

Track: Track 1: Original Research/Position/Education/Attention Track

Keywords: protein language models, AI-biology integration, artificial intelligence safety, biosecurity, dual-use research

TL;DR: Protein language models integrated with wet-lab experiments could accelerate pandemic research but also enable design of novel viruses, demanding new AI safety measures tailored to biological applications.

Abstract: Advances in protein language models (pLMs) and their integration into closed-loop experimental platforms are unlocking powerful new capabilities in protein design. This convergence, termed Intelligent Automated Biology (IAB), enables rapid, large-scale exploration of protein function, accelerating discovery in fields from medicine to synthetic biology. Yet when applied to pathogens, these same tools pose serious dual-use risks. IAB systems can efficiently optimize immune escape, viral fitness, and other dangerous traits, even in the absence of deep biological expertise. In this position paper, we argue that the AI community must take proactive steps to address this emerging AI safety and biosecurity challenge. We introduce a framework categorizing IAB capability levels to guide risk assessment, examine IAB's unique governance challenges, and offer concrete recommendations for pLM-specific safeguard research.

Submission Number: 221