AI Can Design Viruses. Can Our Defenses Adapt Fast Enough?

By Jared Bauer, CEO and Co-Founder, with contributions from Alison O’Mahony, Ph.D. (SVP Research and Development) and Douglas P. Gladue, Ph.D. (VP Virology)

On August 6, 2026, Stanford and Arc Institute researchers published in Science the first peer-reviewed demonstration that AI genome language models can design complete, viable bacteriophage genomes (DOI: 10.1126/science.aec2657). After an earlier version of the article had circulated as a bioRxiv preprint, two Seek Labs scientists, Douglas P. Gladue, Ph.D. and Alison O’Mahony, Ph.D., published a review in Viruses examining how programmable, CRISPR-based antivirals could support adaptable responses to emerging and engineered viruses, including potential threats enabled by advances in AI and synthetic biology (see CRISPR Treatments for AI-Designed Synthetic Viruses: Rapid Programmable Countermeasures for Emerging and Engineered Viruses in Viruses).

What the Stanford Study Showed

The Stanford study, Generative Design of Bacteriophages with Genome Language Models, used the Evo 2 genome language model to generate candidate phage genomes. Evo 2 was trained on 9.3 trillion nucleotides drawn from over 128,000 whole genomes. Researchers chemically synthesized 302 candidate designs, successfully assembled 285, and recovered 16 viable phages capable of infecting and killing E. coli. The designs’ sequence novelty illustrates why surveillance and screening systems that depend heavily on similarity to known viruses may need to evolve.

The study does not show AI can design functional viruses that infect humans. The researchers deliberately focused on bacteriophages and excluded human-, animal-, and plant-infecting viruses from the relevant training data; however, even within those safeguards, the work marks an important transition: a computational model helped produce complete, viable viral genomes with substantial novelty.

For those working at the intersection of biotechnology, AI, and infectious disease, the result adds urgency to longstanding dual-use concerns. In commentary accompanying the Science publication, Tom Inglesby and Mori Hanke of the Johns Hopkins Center for Health Security wrote the result “raises urgent biosafety and biosecurity questions.” The most operationally pressing is whether defensive capabilities can evolve as quickly as design capabilities. That question is central to the work underway at Seek Labs on our novel therapeutic modality, Sequence Ablation Therapeutics (SAT).

Preparing for AI-enabled Biological Threats

The Viruses review published by Seek Labs scientists frames the dual-use dilemma clearly: AI and synthetic biology can accelerate discovery while also increasing the risk of misuse or unintended harm. It examines how programmable CRISPR-based antivirals could support faster, more adaptable countermeasures against emerging and engineered viruses.

Vaccines, monoclonal antibodies, and small-molecule antivirals are generally developed around a defined target; creating a deployable product can take years. Sequence Ablation Therapeutics, which utilize CRISPR, work differently. Once a threat’s sequence is known, researchers can use BioSeeker™ (Seek Labs’ AI System) to rapidly design guide RNAs that target the sequence.

The review proposes a rapid-response pipeline linking real-time sequencing with AI-assisted guide RNA selection, multiplexed cassette design, validation, manufacturing, and potential emergency deployment. Multi-guide-RNA combinations aimed at conserved regions could raise the barrier to viral escape. Cas9-based approaches can target DNA viruses, while Cas13-based approaches can target RNA viruses. Together, these systems offer a family of programmable strategies that could be redirected across viral targets, provided that suitable delivery and safety profiles are established.

Aligning Speed with Oversight

If defense and policy communities take speed seriously, they must confront several hard problems that do not yet have clear solutions.

  • – Rapid target identification requires sequencing capacity near the point of outbreak, which many regions lack.
  • – Rapid therapeutic development requires validated platforms that can pivot to new targets, and those platforms do not yet exist for most pathogen classes.
  • – Manufacturing needs distributed capacity and surge capability, while deployment requires regulatory pathways that can move at emergency speed without compromising safety.

None of these capabilities has been tested together at scale.

The Viruses review also outlines governance measures that could become increasingly important as design tools accelerate, such as pre-cleared guide RNA repositories, transparent design logs, standardized off-target and safety screening, and alignment with evolving nucleic-acid-synthesis screening frameworks. This begins to answer the gap identified by Inglesby and Hanke: biosecurity needs both faster countermeasures and the oversight required to develop and deploy them responsibly.

Keeping Defense Aligned with Biological Design

The Science study is a meaningful scientific advance, and its dual-use implications deserve serious attention: AI-assisted design of complete, viable bacteriophage genomes is no longer hypothetical. Now, the response cannot rely on slow biological design alone; defensive capabilities and oversight must mature alongside it.

That is the work underway at Seek Labs labs today: decoding the sequence of a viral threat, programming a CRISPR response around its vulnerabilities, and resolving it before it spreads—whether the threat emerges naturally or is engineered and whether anyone has seen its exact genome before. With biological design becoming programmable, biological defense needs to be programmable too. That is our vision.

Frequently Asked Questions

Can AI design viruses that infect humans?

Not yet. The Stanford/Arc Institute Science study demonstrated AI-designed bacteriophages that infect bacteria, not people. The researchers deliberately limited the relevant training data and experimental system to avoid human-, animal-, and plant-infecting viruses. Outside biosecurity researchers have argued extending this approach to human disease-causing viruses should not be pursued (The Guardian).

What is programmable CRISPR, and how is it different from a typical antiviral?

Conventional countermeasures are generally developed around defined biological targets. Programmable CRISPR approaches use guide RNAs to recognize specific sequences in a viral genome. Those guides can potentially be redesigned quickly when a target changes, although every revised treatment would still require appropriate validation, delivery, safety, manufacturing, and regulatory work. Seek Labs outlines this approach in a December 2025 Viruses review (DOI: 10.3390/v17121588).

What is a “rapid-response pipeline” for programmable CRISPR countermeasures?

It’s the operational sequence Seek Labs’ Viruses review proposes for responding to a newly identified viral threat: genomic sequencing, AI-assisted guide RNA selection, multiplexed cassette assembly, validation, manufacturing, and potential emergency deployment. The framework also includes governance measures such as pre-cleared guide RNA repositories, transparent design logs, and standardized safety screening.

What did the Science study accomplish?

Researchers used Evo 2, a genome language model, to generate bacteriophage designs, selected 302 candidates, successfully assembled 285, and recovered 16 viable phages. The work is the first peer-reviewed demonstration that generative AI can contribute to the design of complete viral genomes that become living, functioning bacteriophages.

Is Seek Labs arguing AI biology research should stop?

No. AI has real value in the life sciences, including in this study’s own potential applications against drug-resistant bacteria. Seek Labs’ position is that biosecurity defense needs to advance at the same pace as biological design.

What types of viruses can programmable CRISPR countermeasures address?

In principle, a wide range. CRISPR/Cas9 systems can be directed against DNA viruses and CRISPR/Cas13 systems against RNA viruses. Because the guide RNA—not the underlying platform—determines the target, the same modality can be redirected as a virus’s sequence is identified or changes.

Sources

  • – King, S.H. et al. “Generative design of bacteriophages with genome language models.” Science, 2026. DOI: 10.1126/science.aec2657
  • – Arc Institute. “Evo 2: DNA Foundation Model.” NVIDIA model card. build.nvidia.com/arc/evo2-40b/modelcard
  • – King, S.H. et al. “Generative design of novel bacteriophages with genome language models.” Preprint. bioRxiv, September 2025. DOI: 10.1101/2025.09.12.675911
  • – Gladue, D.P.; O’Mahony, A. “CRISPR Treatments for AI-Designed Synthetic Viruses: Rapid Programmable Countermeasures for Emerging and Engineered Viruses.” Viruses 2025, 17, 1588. DOI: 10.3390/v17121588
  • – Sample, Ian. “Safety fears as scientists make first viruses designed by AI.” The Guardian, August 6, 2026. theguardian.com
  • – Thomas V. Inglesby, Moritz S. Hanke, “AI-designed viral genomes.” Science 393,563-564(2026). DOI:10.1126/science.aej8512