AI Has Designed Working Viruses: Why This Matters

Over the past few years, generative AI has learned to produce text, images, audio, video, software code, proteins and molecular designs. Now researchers have demonstrated another step.

An AI model generated complete genetic designs. Scientists selected some of those designs. They were physically produced and tested in a laboratory. Some became functioning bacteriophages.

That makes this different from AI merely predicting biology on a screen. The important milestone is not that AI created a human pathogen. It is that AI-generated biological designs crossed from software output into functioning biology.

The Quick Answer

What Scientists Actually Created

The researchers did not ask a chatbot: 'Make a virus.' They used specialised biological AI models trained to understand patterns in genetic information.

Researchers working with models including Evo 1 and Evo 2 — genome language models developed at the Arc Institute in collaboration with researchers from Stanford University and the University of California, Berkeley — trained these systems on large quantities of bacteriophage genetic information. Critically, viral genomes associated with humans, animals and plants were deliberately excluded from the training data.

The systems generated thousands of candidate bacteriophage genome designs. Researchers then:

• Assessed candidate designs
• Selected a subset for physical testing
• Used controlled laboratory processes to produce the selected candidates
• Checked whether viable bacteriophages resulted
• Tested whether those bacteriophages could infect target bacteria

Researchers reported testing nearly 300 candidate genomes. Of those, 16 produced viable bacteriophages. Some AI-designed phages demonstrated the ability to function against Escherichia coli. Researchers also reported that mixtures of AI-designed phages were capable of overcoming resistance in bacterial strains.

The study was published in Science, one of the world's leading peer-reviewed scientific journals. Verify current detail on the journal's official website and the researchers' institutional pages.

What a Bacteriophage Is

A bacteriophage — often shortened to phage — is a virus that infects bacteria. It cannot infect human cells in the way that viruses such as influenza or coronaviruses do.

How AI Was Used

Evo 1 and Evo 2 are genome language models — AI systems that learn statistical patterns in genetic sequences, in a similar manner to how large language models learn patterns in text. Instead of predicting the next word in a sentence, these models can generate the next section of a genome by learning from millions of existing genetic sequences.

The models were trained specifically on bacteriophage genetic information, with human, animal and plant viral sequences excluded. This deliberate constraint was part of the researchers' approach to keeping the work focused on bacteriophages and avoiding training the models on the kinds of viruses most likely to cause harm to people.

The models generated many thousands of candidate genome designs. Most of these were not selected for physical testing. Those that were selected went through a physical laboratory process in order to produce actual bacteriophages for testing.

What the Experiment Demonstrated

The study demonstrated that:

• Generative AI can produce some complete viral genome designs
• A subset of generated designs can be biologically viable when physically produced and tested
• AI may allow researchers to explore biological design spaces that humans would struggle to examine manually at this scale
• AI-designed phages may show useful functional properties
• AI-designed phages may eventually contribute to therapeutic research, particularly regarding antibiotic-resistant bacteria

What the Experiment Did Not Demonstrate

It is equally important to be clear about what the experiment did not demonstrate:

• AI can easily design every type of virus
• AI can readily produce human pathogens
• Human scientific expertise is no longer necessary
• The laboratory and physical manufacturing bottlenecks have disappeared
• Every generated design works — the majority of tested candidates did not produce viable phages
• Biological synthesis is now automatic
• Existing biosecurity countermeasures are obsolete
• Unrestricted AI access inevitably leads to biological weapons

The failure rate was significant. Researchers reported testing nearly 300 candidate genomes, of which 16 produced viable bacteriophages. That is a small proportion of those tested, and those tested were a small selection from thousands generated. Biological complexity remains a real barrier.

Bacteriophage genomes are also relatively small and compact compared with the genomes of many other biological systems. The challenge of designing larger, more complex biological systems would be substantially greater.

Why This Could Matter for Medicine

Bacteriophages are being researched and used in some settings as potential treatments for bacterial infections. Interest has grown considerably because of antibiotic resistance — the increasing ability of bacteria to survive medicines that would otherwise control them.

Potential advantages of phage therapy under research include:

• Targeting specific bacteria with precision
• Attacking antibiotic-resistant strains through different biological mechanisms
• Evolving alongside bacterial populations
• Complementing rather than replacing existing therapies

But phage therapy also faces significant challenges:

• Narrow target range — a phage effective against one bacterial strain may not work against others
• Bacteria can develop resistance to phages
• Immune responses may limit effectiveness
• Manufacturing and standardisation at scale
• Regulatory approval processes
• Clinical evidence requirements

Do not rely on this article for medical advice. Consult qualified healthcare professionals for decisions about treatment.

The Antibiotic Resistance Opportunity

Antibiotic resistance is a significant global health problem. Bacteria evolve, and some strains have developed the ability to survive medicines that previously controlled them reliably. In some cases, very few effective options remain.

Phage therapy may offer an additional route because bacteriophages can target bacteria through different biological mechanisms than antibiotics. AI could potentially help researchers:

• Explore larger biological design spaces than manual research can examine
• Identify candidate phages more efficiently
• Adapt designs to particular bacterial resistance patterns
• Shorten parts of the early research cycle

Why Scientists Are Raising Safety Concerns

The central concern is not that this particular experiment created a dangerous pandemic pathogen. It did not. The concern is about what the experiment demonstrates in principle.

Once an AI system has demonstrated the ability to generate functional biological designs, the same class of technology may eventually become relevant to:

• Beneficial medicine and phage therapy
• Industrial biotechnology
• Agriculture and food production
• Environmental science
• Harmful biological design

Experts commenting on the research have emphasised that the immediate risk from this study remains constrained by substantial barriers including:

• Deep scientific expertise required
• Biological complexity of more dangerous systems
• Physical synthesis — digital designs must be manufactured
• Specialist laboratory capability
• Testing and evaluation
• Containment requirements

These barriers matter now. But they may change over time as AI models improve and as access to biological infrastructure becomes more widely distributed.

Biosafety Versus Biosecurity

Most laboratory work focuses primarily on biosafety — the controls that prevent accidents. Biosecurity is a separate and sometimes harder problem: it is about deliberate misuse rather than accidental harm, and it requires thinking not only about the researchers in a specific institution but about who else might eventually access similar capabilities.

The Dual-Use Problem

Almost all biological research is dual use to some degree. Understanding how bacteria develop resistance to antibiotics is necessary for fighting infection — but the same knowledge could theoretically be applied to make infections harder to treat.

AI-assisted genome design intensifies this long-standing challenge. The same model that helps researchers design bacteriophages for therapeutic research might theoretically be guided towards less benign goals.

The Digital-to-Physical Boundary

Most generative AI risk discussions focus on information. Examples include misinformation, phishing emails, software code, persuasive text and copyrighted material. The harm happens in the information domain.

Biological AI adds a different dimension. A digital design can potentially become:

• A DNA sequence
• A protein
• A molecule
• A biological system

This transition requires physical infrastructure, specialist knowledge and laboratory controls. The physical world is the check at the moment. But as AI improves and biological infrastructure becomes more accessible, the significance of that physical checkpoint changes.

Why DNA Synthesis Is a Critical Control Point

Experts in biosecurity and responsible research increasingly focus on DNA synthesis providers. These companies sit at an important junction between a digital design and a physical biological object.

Responsible synthesis providers use screening systems intended to identify requests that may relate to controlled, dangerous or suspicious biological material. These systems involve assessing the sequence, the customer, the stated use and the applicable legal requirements.

This article does not describe how synthesis screening works at a technical level, or how it might be circumvented. The point is that responsible providers recognise their role in the governance chain — and that role is significant.

Why AI Controls Alone Are Not Enough

It is tempting to think that the solution is simply to restrict access to AI models — or to add safety filters to the models themselves. These controls matter, but they are not sufficient on their own.

Biological AI risk is distributed across the full chain from model training to physical production. A model with good safety controls does not prevent somebody from using a different model, accessing scientific literature, using earlier open-source models or finding other routes to similar information.

Layered Governance

Responsible approaches to AI-assisted biological research require governance at multiple levels simultaneously:

These layers do not replace each other. Weakness in one layer places greater pressure on the others. A model with strong access controls does not remove the need for laboratory biosecurity, and laboratory controls do not eliminate the need for thoughtful governance of who can access the models in the first place.

What This Means for Businesses

Most businesses will never design biological systems. This research is not a direct operational risk for the vast majority of organisations.

But the experiment illustrates five governance lessons that apply broadly:

1. 1AI CAPABILITIES ARE EXPANDING RAPIDLY — Policies written around text-generating chatbots may become outdated quickly as the same underlying technology extends into new domains.
2. 2OUTPUT RISK DEPENDS ON CONSEQUENCE — Generating an email and generating a biological design require radically different controls. Governance should be calibrated to what the output can cause, not just what the tool looks like.
3. 3HUMAN OVERSIGHT REMAINS ESSENTIAL — The researchers selected, tested and evaluated AI outputs at every stage. The AI generated candidates; humans made all the decisions about what to do with them.
4. 4DIGITAL CONTROLS MAY NOT BE ENOUGH — Physical-world systems and suppliers may also need safeguards. The synthesis step in this research is one example; payment processing, physical access and supply chains provide others.
5. 5GOVERNANCE MUST CHANGE WITH CAPABILITY — A tool approved for one purpose should not automatically remain approved when its capabilities expand or when it is connected to new systems.

What This Means for AI Governance

AI governance frameworks developed around text generation and document automation are adequate for many ordinary business uses. But as AI demonstrates the ability to influence physical systems — whether through biological design, autonomous agents, physical robotics or infrastructure control — those frameworks require revision.

Powerful AI should be assessed according to:

• Capability — what can it actually do?
• Access — who can use it and under what conditions?
• Data — what information does it receive?
• Tools — what can it call or trigger?
• Downstream actions — what can happen as a result of its outputs?
• Irreversible consequences — what cannot be undone?
• Human approval — where must a person decide before action is taken?
• Monitoring — is activity observable and recorded?
• Specialist oversight — does this domain require expert review beyond standard IT governance?

The IT Club AI Governance hub covers these principles in more depth, including guidance on AI risk assessment, human review, supplier assessment, approved tools and incident response.

Explore the IT Club AI Governance Knowledge Centre →

How to Carry Out a Business AI Risk Assessment →

High-Consequence AI: When Outputs Affect the Physical World →

What This Means for Scientific Research

The research community is already engaged in active debate about how to handle AI-assisted biological design responsibly. Questions under active discussion include:

• Whether and how to publish dual-use research findings
• What pre-publication review is appropriate for AI-assisted genome research
• How to design safety evaluations for biological AI models
• How to govern access to powerful biological AI systems
• Whether existing international agreements adequately cover AI-assisted design
• How research institutions should update their oversight frameworks
• How synthesis providers should approach AI-generated sequences
• Who bears responsibility for downstream misuse of published research

These are not new questions — biological research has grappled with dual-use concerns for decades. But AI-assisted design may accelerate the pace at which capabilities become more broadly accessible, making the governance questions more urgent.

Warning Signs

Higher-risk AI deployments deserve specialist review where any of the following apply:

• AI output can directly create or influence physical outcomes
• Research involves biological, chemical or physical systems
• Generated material could affect public safety
• Irreversible actions are automated without human approval
• Specialist scientific or technical judgement is absent from the review process
• Tool capability changes without reassessment
• External suppliers form part of the safety boundary
• Audit trails are inadequate or absent
• Access to powerful AI tools is unnecessarily broad
• Dangerous or unusual capabilities have not been systematically tested
• Incidents cannot be quickly contained
• Responsibility for output is unclear

Practical Business Implications

AI has moved beyond content generation

Generative systems now produce designs for biological systems, molecules, materials and software components that can be physically produced. Governance frameworks should reflect this expanding capability rather than treating all AI as text generation.

The research is about bacteriophages

It did not demonstrate that AI can readily design human pandemic viruses. The distinction matters, and sensational headlines do not help organisations make proportionate decisions.

The medical potential is significant

AI-assisted phage design may eventually contribute to research against antibiotic-resistant bacteria. That benefit is worth supporting and not worth discarding because of governance concerns that can be addressed with appropriate controls.

The failure rate was high

Researchers reported that only 16 of nearly 300 tested AI-generated genomes produced viable bacteriophages. Most did not work. Biological complexity remains a real constraint on AI-assisted design.

Physical synthesis remains important

Digital AI capability does not remove the physical manufacturing stage. Responsible synthesis providers, laboratory controls and specialist expertise are all still required in the chain from design to biological object.

Biosecurity needs several layers

AI-model controls alone are not sufficient. Synthesis screening, laboratory containment, research ethics oversight, institutional review and international coordination all form part of an adequate governance response.

Governance must follow capability

As AI demonstrates the ability to produce more consequential designs, the controls around who can use it, what they can do with it and how outputs are reviewed must evolve to match.

The IT Club View

This is a significant scientific achievement. AI-designed bacteriophages may eventually contribute to better therapies, biotechnology tools and research approaches. Dismissing the medical potential would be as unhelpful as treating the experiment as evidence that AI poses an imminent pandemic threat.

But the experiment also demonstrates something that regulators and businesses should notice: AI is increasingly capable of producing designs that can be converted into real-world systems.

IT Club's view is that the appropriate response is to:

• Continue beneficial research — AI-assisted biological design has real medical and scientific value
• Avoid sensationalism — accurate framing helps organisations make proportionate decisions
• Evaluate dangerous capabilities systematically before releasing powerful models
• Involve specialist safety professionals in research review at the frontier
• Keep humans accountable for decisions made using AI designs
• Control access proportionately — higher-risk capabilities warrant more restricted access
• Maintain synthesis screening — the physical manufacturing step is a critical control point
• Maintain laboratory biosafety and biosecurity standards
• Use layered governance — no single control is sufficient on its own
• Review controls as capabilities evolve — governance written for yesterday's AI may be inadequate for today's

The objective should not be to stop AI-driven biological research. It should be to make sure scientific capability grows alongside the controls needed to use it responsibly.

Plain-English Takeaway

Researchers have demonstrated that generative AI can design complete bacteriophage genomes that become functioning viruses when produced and tested in a laboratory. These viruses target bacteria rather than humans, and the research could eventually help areas such as phage therapy and antibiotic resistance. The wider significance is that AI-generated biological designs can now cross from digital output into functioning biology, increasing the importance of model safeguards, responsible research, DNA-synthesis screening and laboratory biosecurity.

Administrator / Governance Technical Note

Operational Heartbeat

Advanced AI capability changes as models improve, new domains become accessible, tool access expands, output quality increases and physical-world integrations grow. Regulations change. Supplier safeguards change. Research reveals new capabilities and new risks.

A recurring high-consequence AI review should check:

• Approved purpose — is the AI still being used for what it was assessed for?
• Capability — has the model's capability changed since last review?
• Users — who has access, and is that still appropriate?
• Access — are permissions still correctly calibrated?
• Data — what information is entering the system?
• External tools — what can the AI call or trigger?
• Downstream actions — what can happen as a result of outputs?
• Human approval — are the right decisions still requiring human sign-off?
• Supplier safeguards — are the provider's controls still adequate?
• Incidents — have there been near-misses, errors or complaints?
• Research developments — has new work changed the risk picture?
• Regulatory changes — do new laws or guidance apply?
• Specialist review — has appropriate expert input been obtained?
• Corrective actions — are outstanding issues being addressed?
• Next review date — has a date been set and owner assigned?

Related Business Questions

What does this kind of AI capability mean for ordinary businesses?

Ask the IT Club Advisor about AI governance, high-consequence automation, human approval, risk assessment or how to decide which AI uses need stronger controls.

Ask Your IT Question →

Free to ask. No credit card. No sales pressure. Fair usage applies.