Preparing for Future AI Risks in Biology

AI models hold powerful potential to accelerate biological discovery, such as drug development, vaccine design, and enzyme creation. However, they simultaneously carry a dual-use problem, where the ability to reason about biological data or predict chemical reactions could be misused for malicious purposes such as bioweapon creation.

OpenAI expects future models to reach 'High' level capability in biology, and is pursuing a multi-layered approach to prevent this. The key strategies are as follows:

• Developing a responsible approach to biological capability advancement
• Collaborating with external experts such as government agencies and national laboratories
• Training models to safely handle dual-use biological requests
• Building detection, monitoring, and enforcement systems
• Deploying adversarial red-teaming tests and security controls through experts

OpenAI prioritizes prevention above all, and is collaborating with organizations such as US AISI, UK AISI, and Los Alamos National Lab to advance biosecurity threat models and capability evaluations. It is also building practical defense systems by conducting red-teaming tests involving experts and researching the role of AI in laboratory settings.