AI Drives Autonomous Biological Design: The Dual-Use Challenge

Artificial intelligence (AI) is demonstrating the capability to autonomously design and execute biological experiments, with governance systems struggling to keep pace.

In February 2026, OpenAI and Ginkgo Bioworks announced a significant milestone: OpenAI's GPT-5 model had autonomously designed and run 36,000 biological experiments via a robotic cloud laboratory. This groundbreaking process involved the AI proposing study designs, robots executing them, and data being fed back to the model for iteration.

Humans established the overarching goal, while machines performed most of the intricate lab work, resulting in a 40% cost reduction for producing a specific protein.

This process involved the AI proposing study designs, robots executing them, and data being fed back to the model for iteration, resulting in a 40% cost reduction for producing a specific protein.

The Dawn of "Programmable Biology"

This development, termed "programmable biology," involves designing biological components digitally and constructing them physically, with AI completing the experimental loop. This marks a significant shift from traditional biology, which primarily focused on observation and understanding, evolving from genome sequencing to advanced gene editing tools like CRISPR.

AI now accelerates a third phase where computers design and rapidly test biological systems. This innovative approach closely resembles engineering principles, emphasizing a continuous cycle of design, build, test, learn, and repeat. This allows for the parallel exploration of thousands of design variations, drastically speeding up discovery and development.

AI Accelerates Protein Design

AI-accelerated protein design stands out as a prime example of this new automation. Proteins are the workhorses of living cells, performing most functions, yet designing new ones has historically been a process of extensive trial and error.

Now, advanced protein language models, trained on millions of natural protein sequences, can predict the effects of mutations or design entirely novel proteins. These models are already contributing to new drug development and vaccine acceleration. When combined with automated laboratories, these powerful models create rapid experimental and revision cycles, completing tasks in days that would traditionally take months or even years.

The Dual-Use Dilemma

This powerful technology presents a significant "dual-use problem," meaning tools developed for beneficial purposes could inadvertently be misused.

Researchers have found that AI models integrated with automated labs can optimize viral spread, even without specialized training.

Moreover, risk-scoring tools have been developed to evaluate how AI could modify viral capabilities, such as altering host species or evading immune systems. Crucially, AI models are also capable of guiding users through technical steps to recover live viruses from synthetic DNA, potentially lowering barriers to bioweapon development. This particular risk, experts note, is not adequately addressed by current oversight mechanisms.

Novice Biosecurity: Mixed Findings

Research exploring whether AI can empower individuals with limited biology training to conduct dangerous lab work has yielded mixed conclusions.

A study conducted by Scale AI and SecureBio indicated that novices using large language models could complete biosecurity-related tasks with four times greater accuracy and sometimes even outperformed trained experts. Alarmingly, approximately 90% of these novices reported minimal difficulty obtaining risky biological information despite existing safety filters.

Conversely, a separate study led by Active Site found that AI assistance did not significantly improve novices' overall ability to complete complex workflows required to produce a virus in a biosafety laboratory. However, the AI-assisted group did demonstrate higher success rates and faster completion times for specific steps, such as cell growth.

Regulatory Landscape Lags Behind

The traditional reliance on skilled human hands for laboratory work is rapidly decreasing. Cloud laboratories and robotic automation are becoming increasingly accessible, enabling AI-generated experimental designs to be executed remotely with ease.

Existing regulations for biological research do not account for AI-driven automation, and AI regulations do not specifically address its biological applications.

The regulatory vacuum is evident:

• In the U.S., a 2023 executive order on AI security that included biosecurity provisions was later revoked.
• Screening of synthetic DNA by commercial providers remains largely voluntary.
• A 2026 bipartisan bill proposing to mandate DNA screening does not yet address AI-designed sequences that might evade detection.
• The 1975 Biological Weapons Convention, a cornerstone of international biosecurity, entirely lacks provisions for AI.

Addressing the Risks: Proposals and Challenges

Safety evaluations conducted by AI labs themselves are often criticized as opaque and insufficient to adequately assess real-world risks. Researchers estimate that even minor improvements in AI's ability to plan pathogen-related experiments could lead to thousands of additional deaths from bioterrorism annually.

To mitigate these escalating risks, several proposals have been put forward:

• A managed access framework for biological AI tools.
• Improved DNA synthesis screening.
• Rigorous model evaluations before AI systems are released.
• Enhanced governance of biological data, particularly genomic data.

While some AI companies have voluntarily implemented safety measures, their leaders acknowledge a critical challenge: the pace of AI development may soon exceed individual company assessment capabilities, necessitating broader, coordinated efforts.

The Unanswered Policy Question

While AI clearly demonstrates its ability to facilitate research goals in controlled environments, the profound implications of these capabilities operating outside such controls remain an urgent and unanswered policy question.

Overreaction risks diverting talent and investment, while under-reaction could allow exploitation of the technology for harm.

This final thought underscores the delicate balance required in navigating this transformative technology.