Safeguarding the Future of AI in Biology: Balancing Innovation and Biosecurity

Preparing for Future AI Capabilities in Biology

AI models are advancing their abilities in biology, offering new opportunities and challenges. These models are already assisting scientists in identifying promising drug candidates for human trials and hold potential to accelerate drug discovery, improve vaccine design, create enzymes for sustainable fuels, and develop treatments for rare diseases.

However, these capabilities also raise dual-use concerns. The same tools that support scientific progress could be misused to recreate biological threats or aid in bioweapon development. While physical access to labs and sensitive materials remains a barrier, it is not absolute. Upcoming AI models are expected to reach high levels of biological capability, prompting a multi-faceted approach to risk mitigation.

Our Approach

Responsible action is necessary given the uncertainty around these developments. The focus is on enabling positive applications like biomedical research and biodefense while restricting harmful uses. Early intervention is preferred over reactive measures after a bio threat event occurs.

Collaboration with external experts has shaped the biosecurity threat model, capability assessments, and usage policies. Human trainers with advanced biology degrees have contributed to evaluation data, while domain-expert red teams test safeguards under realistic conditions. Partnerships with government agencies and national labs support biosecurity tools and assessments.

Capability assessments, informed by expert input, estimate when AI models reach high-risk thresholds. Although these assessments cannot predict real-world misuse with certainty, preparedness and readiness measures are prioritized due to the stakes involved.

Strengthening Defenses in Biology

Over the past two years, the development of frontier AI models has included ongoing risk tracking, risk reduction before launch, and transparent sharing of findings through system documentation. Preparedness evaluations run during model training provide early indicators of capability levels.

Training AI to Refuse or Safely Respond to Harmful Requests

Models are trained to refuse requests that are explicitly harmful or related to bioweaponization. For dual-use requests (like virology experiments or genetic engineering), responses avoid providing actionable steps. The goal is to offer high-level insights supporting expert understanding without enabling misuse by novices.

Always-On Detection Systems

System-wide monitors detect risky bio-related activity across all product interfaces. Unsafe requests are blocked, triggering automated and human reviews when necessary.

Monitoring and Enforcement Checks

Use of AI products to cause harm is prohibited. Advanced AI reasoning assists in detecting biological misuse, supported by human reviewers. Consequences for misuse include account suspension and potential law enforcement notification in severe cases.

End-to-End Red Teaming

Expert red teams attempt to bypass safety measures, simulating well-resourced adversaries. Given the challenges in biology domain expertise and AI exploitation skills, teams pair AI security experts with biological domain experts to thoroughly test safeguards.

Security Controls

A defense-in-depth strategy protects model weights through access controls, infrastructure hardening, egress controls, and monitoring. Detection and response systems, threat intelligence, and insider-risk programs operate continuously to identify and block threats. These measures have been reviewed by a Safety and Security Committee and are applied to current models that remain below high-risk capability levels.

What’s Ahead

The broader challenge includes the possibility of widely accessible AI bio capabilities combined with increasingly available life-sciences synthesis tools. To address this, a biodefense summit is planned, gathering government researchers and NGOs to discuss dual-use risks, share progress, and explore AI’s role in accelerating research.

Efforts aim to deepen partnerships with governments and enhance biodefense work, including countermeasures and novel therapies. There is also a focus on developing protocols to grant vetted institutions access to the most capable models to advance biological sciences, including diagnostics and testing methods.

Strengthening society’s defenses against biological threats requires collaboration between public and private sectors. This may involve improved screening of nucleic acid synthesis, enhanced early detection systems for pathogens, infrastructure hardening, and investment in biosecurity innovations.

The intersection of AI and biosecurity will likely create opportunities for mission-driven startups and investment in safety and security as essential services. Ongoing collaboration with governments, researchers, and entrepreneurs will be critical to prepare the biosecurity ecosystem while leveraging scientific breakthroughs.