A Critical Moment for AIxBio Safety and Security

The convergence of artificial intelligence and the life sciences is reshaping what’s possible in biology, delivering major advances–ranging from accelerating vaccine discovery to advancing new tools for mitigating the effects of climate change–and new risks. NTI | bio’s AIxBio Horizon Scan looks ahead to where AIxBio capabilities are headed over the coming years. This Horizon Scan is the first in a series that we plan to conduct every six months.

Our horizon scan supports international efforts to understand emerging AI risks and is intended to help researchers, industry leaders, and policymakers make well-informed decisions about how to get out ahead of those risks. Researchers can use it to design effective safeguards, industry practitioners to inform responsible AI development, and government officials to craft policies that keep pace with rapid change.

Anticipating What’s Next

AIxBio capabilities are advancing faster than many experts expected. For example, large language models now outperform human virologists in answering expert-level questions, and natural language interfaces are putting powerful biological design tools into the hands of users without deep technical training.

This creates a fundamental challenge: how can policy makers, model developers, and researchers make sound decisions about safety, security, and governance when capabilities evolve so quickly?

Our horizon scan helps address this challenge by tracking emerging capabilities and identifying trends before they fully materialize. Think of it like a weather forecast—showing what conditions are developing and what’s likely ahead. Risk assessments can then build on this forecast to identify specific vulnerabilities and potential high-leverage intervention points to enhance biosecurity.

Both are essential. But without a forecast, even the best-designed biosecurity interventions may target yesterday’s problems rather than tomorrow’s realities.

What We’re Seeing—and Why It Matters

The scan highlights four key trends in AIxBio development:

1. Integration of LLMs with biological research tools: AI agents use an LLM as its “brain,” combined with a memory system and task-specific tools. Agents equipped with biological AI tools and laboratory robotics could enable testing and optimization of AI-designed molecules alongside broader biological discovery.
2. Automation and laboratory integration: Automation will be critical for AIxBio progress. Cloud labs and automated platforms are dramatically lowering barriers to experimentation, removing the need for years of hands-on training, and they may enable fully automated AI-driven research without humans in the loop within a decade. Systems will therefore require deliberately designed safeguards rather than relying on humans to flag potentially harmful work.
3. Data as a bottleneck: Data collection is a significant bottleneck for developing more powerful biological AI models that make higher fidelity predictions. Large amounts of additional data will need to be collected to significantly advance these models beyond current capabilities.
4. Progress at different levels of biological scale: At the molecular level, biophysics-based AI models show promise for drug discovery and may enable capabilities beyond current training data, such as designing enzymes with novel functions. At the cellular level, “virtual cells” are emerging as potentially powerful predictive tools.

Over the next 6 to 18 months, AI agents will better integrate with biological research tools, automating tasks that currently require specialized expertise–like analyzing genomic sequences or identifying patterns in complex biological datasets. This could allow small teams to achieve outcomes that once required large, well‑funded labs.

AI-enabled automated laboratory platforms will proliferate, moving toward facilities with minimal human involvement. This shift will challenge existing oversight mechanisms which rely on review by trained scientists, and it raises questions about where safety checks should be built into fully automated design‑build‑test cycles.

This timeframe will also see the emergence of “closed-loop discovery systems” capable of automatically synthesizing, testing, and refining AI-designed proteins or genomes. Rather than human scientists interpreting results and designing follow-up experiments, these systems will autonomously iterate through hundreds or thousands of experiments. The implications cut both ways: unprecedented speed for beneficial discoveries, but also the potential for dual-use tools to be developed without meaningful opportunity for risk reduction interventions.

Looking ahead 2 to 5 years, LLMs will likely be able to orchestrate biological workflows at or beyond expert human levels. We may see the emergence of biological programming languages where researchers specify desired biological functions, and AI systems design and produce the corresponding genetic circuits, proteins, or organisms.

A Critical Moment for Action

The AIxBio field stands at a critical juncture. Capabilities are advancing rapidly, but governance frameworks and safety measures lag behind. The next 18 months may determine whether safety practices, emerging evaluation frameworks, and national and international governance frameworks can keep pace with technological change.

The window for establishing robust governance to safeguard rapidly advancing new technologies is often narrow. This horizon scan helps decision-makers take action within that window–proactively reducing risks before it closes.