When AI enters the laboratory: technological Dual-Use changes scale

Artificial intelligence is gradually crossing an important frontier. After learning to generate text, images, computer code, molecular structures, and proteins, it is now beginning to take part in the design of far more complex biological systems. In August 2026, researchers from Stanford and theArc Institute published work in Science demonstrating that artificial intelligence models could generate complete genomes of functional bacteriophages, that is, viruses that infect bacteria. The researchers obtained 16 viable phages with different biological profiles, and some demonstrated an interesting ability to fight bacteria that had become resistant to the natural phage used as a reference.

This advance opens up considerable possibilities. Bacteriophages have long been the subject of research as a potential solution or complement to antibiotics against certain bacterial infections, particularly in a context where antimicrobial resistance represents a major challenge. The possibility of using AI to explore biological space more quickly and to design new variants could eventually accelerate that research.

It also highlights a fundamental property of many advanced technologies: the same capability that makes it possible to solve one problem can sometimes increase the ability to create another. Artificial intelligence accelerates research, expands the space of possibilities, and lowers certain barriers to experimentation. These characteristics can produce remarkable advances in medicine, pharmaceuticals, and biotechnology. At the same time, they compel organizations to think much earlier about the possible consequences of the capabilities they develop.

The real challenge therefore goes well beyond bacteriophages. As artificial intelligence enters laboratories and takes part directly in biological design, technological Dual-Use changes scale. Governance will have to evolve along with it.

The work published in Science is an excellent example of this new reality. The researchers used genomic models capable of learning the structures and constraints present in large quantities of DNA sequences. Evo 2, presented earlier in 2026 in Nature, had been trained on roughly nine trillion base pairs drawn from every domain of life. The objective is to allow artificial intelligence to work directly with the fundamental language of biology.

For the bacteriophage experiment, the virus ΦX174, which infects E. coli, served as the starting model. Hundreds of AI-generated genomes were selected, synthesized, and tested in the laboratory. Sixteen produced viable phages. Some displayed characteristics different from the natural virus, and a mixture of several generated phages succeeded in infecting bacteria that had developed resistance to the natural reference phage.

It is important to measure what this experiment represents without making it say more than it demonstrates. This is not the creation of a human virus by artificial intelligence. It is the demonstration that a generative model can help design complete, functional viral genomes. That capability is significant enough that biosecurity researchers immediately underlined its potential implications. A perspective published alongside it in Science argues that the generation of functional viral genomes raises urgent questions of biosafety and biosecurity.

This is precisely where the notion of Dual-Use becomes essential. A Dual-Use technology has beneficial applications while also being capable of producing other uses or consequences that carry risk. This principle has long existed in scientific research, in cybersecurity, in the nuclear field, in aeronautics, in space technologies, and in many engineering disciplines. Artificial intelligence, however, adds a new dimension: it can considerably accelerate the speed of exploration and make certain capabilities accessible to more people and more organizations.

In the biological field, this acceleration can be extraordinarily beneficial. Bacteriophages have been studied as a therapeutic avenue for more than a century and are seeing renewed interest today in the face of hard-to-treat infections and antibiotic resistance. The possibility of using models to identify, design, or optimize phages more quickly could help explore a far larger number of candidates and eventually develop more precise therapeutic approaches.

Artificial intelligence can also accelerate protein design, genomic analysis, molecule discovery, the understanding of biological interactions, and many other activities that previously required much longer experimental cycles. Part of the space of biological possibilities can be explored digitally before selecting the candidates that merit physical experimentation.

This capability, however, changes the governance question. When a technology makes it possible to explore ten possibilities, a human team can examine each one relatively easily. When it makes it possible to explore ten thousand or a million, the problem becomes different. Generative capacity grows much faster than the human capacity to individually examine every possible consequence.

Here we find a phenomenon already visible in other areas of artificial intelligence: scale changes the nature of the risk.

In a traditional laboratory, governance rests in particular on people, equipment, substances, protocols, authorizations, and the experiments actually carried out. With artificial intelligence, a significant part of the exploration can begin well before the physical experiment. Models can generate hypotheses, propose structures, and explore combinations in a purely computational environment.

The frontier of governance must therefore also move upstream.

This is where a principle of Dual-Use by Design becomes interesting. It consists of considering both the beneficial possibilities and the potentially problematic uses from the moment a capability is designed, rather than waiting for its deployment to start thinking about the necessary control mechanisms.

This approach does not mean that a technology with Dual-Use potential should be avoided. A great many of the most useful technologies in our history have precisely this characteristic. It means instead that risk management becomes a property of the architecture and of the development process.

For a pharmaceutical, biotechnology, or research organization, this could begin with a simple question: what capabilities are we actually creating?

This question goes further than an inventory of the models in use. A relatively general model can acquire very different capabilities when it is connected to specialized databases, scientific tools, computing infrastructure, automated systems, or laboratory equipment. Risk must therefore be assessed at the level of the complete system and its interactions, not solely from the model taken in isolation.

Identity and permissions also take on a new dimension. Who can use a model? What data can it access? What tools can it call? What categories of experiments can it propose? Can it trigger an action automatically? Which results require additional validation? Which activities must remain under explicit human authorization?

These questions closely resemble those organizations are already beginning to encounter with AI agents in their IT environments. The difference lies in the nature of the possible consequences. When a system acts on code, data, or digital infrastructure, its radius of action must be controlled. When it takes part in a chain that could eventually lead to a biological or physical realization, governance must also consider that passage from the digital to the real.

The concept of least privilege then becomes particularly relevant. A system should have only the data, tools, and capabilities necessary for the task entrusted to it. Sensitive environments can be segmented. Certain actions can require several levels of authorization. Activities can be logged and results made traceable enough to make it possible to understand which data, which models, and which human interventions contributed to the process.

This traceability becomes fundamental when several artificial intelligences take part in the research. One AI can propose a hypothesis, a second analyze it, a third run a simulation, and a fourth help prepare an experiment. When the final result carries significant consequences, the organization must be able to reconstruct that chain of reasoning and action.

Dual-Use by Design thus connects with several principles we already know in cybersecurity: separation of duties, least privilege, segmentation, logging, access control, independent validation, oversight, and the ability to interrupt an operation. These principles can be adapted to environments where artificial intelligence, scientific research, and physical systems converge.

The objective is to enable innovation with a level of control proportional to the capabilities being developed.

This approach becomes all the more important as AI helps reduce the cost of exploration. In several scientific fields, a significant part of the difficulty traditionally lay in the limited number of hypotheses a team could reasonably examine. Generative systems can multiply those possibilities. What was once improbable simply because no one had the time to explore it can become accessible to a machine capable of traversing a considerable space of solutions.

That power is precisely one of the reasons artificial intelligence can accelerate scientific discovery. It is also the reason governance mechanisms must evolve at the same pace as the capabilities.

The solution is not to systematically slow down innovation. In fields such as antimicrobial resistance, drug development, or understanding disease, accelerating research can produce considerable human benefits. Recent progress in phage therapy illustrates a field where new approaches could address hard-to-treat infections.

True maturity consists of developing, at the same time, the capacity to innovate and the capacity to govern that innovation.

This also requires collaboration across disciplines. Artificial intelligence specialists cannot, on their own, determine all the biological consequences of a system. Biologists should not have to become specialists in security architecture, IAM, or agent governance. Cybersecurity teams must understand that some scientific infrastructure now combines data, models, automation, and physical systems in a way that goes beyond the traditional boundaries of IT.

Dual-Use by Design governance therefore necessarily becomes multidisciplinary. It brings together the scientists who understand the field, the AI specialists who understand the models, the security teams who understand the control mechanisms, the compliance officers who understand the obligations, and the executives who must determine the level of risk acceptable for the organization.

This convergence is particularly important for the pharmaceutical, biotechnology, medical, government, and defense sectors. These organizations can benefit enormously from the capabilities of artificial intelligence while working with knowledge, infrastructure, and research whose consequences sometimes extend far beyond their own IT environment.

Hypersecurity takes on its full meaning here. Security no longer concerns only the protection of a server, a database, or a user account. It must consider human and machine identities, models, scientific data, the tools accessible to agents, computing infrastructure, suppliers, physical equipment, validation chains, and the possible consequences of an action.

When the digital can directly influence the biological, the frontier of cybersecurity expands.

The bacteriophages generated with the help of artificial intelligence models represent a remarkable scientific advance. They show that AI can begin to intervene in the design of complete biological systems and that it could eventually help accelerate fields such as phage therapy and the fight against certain antibiotic-resistant infections.

At the same time, this advance teaches us something much broader about the evolution of artificial intelligence. As its capabilities leave the screen to reach laboratories, equipment, infrastructure, and the physical world, the consequences of its decisions and its creations also change scale.

Technological Dual-Use then becomes an architectural reality. A single capability can open a new therapeutic path, accelerate a discovery, and make it possible to explore possibilities that were previously out of reach, while requiring governance mechanisms proportional to the power it makes available.

For Quantum Beyond, this evolution reinforces the importance of an approach in which innovation, architecture, AI Governance, and Hypersecurity advance together. Organizations working at the frontier of AI, pharmaceuticals, biotechnology, healthcare, research, or defense will need specialists in their respective fields. They will also need a cross-cutting view capable of understanding how models, data, identities, permissions, infrastructure, and validation processes interact.

Dual-Use by Design can become one of the principles of this new generation of governance: identifying capabilities as they appear, understanding their possible uses, limiting privileges where necessary, preserving traceability, maintaining human validation proportional to the consequences, and designing security mechanisms at the same time as the innovation.

AI will probably allow us to explore scientific possibilities we would never have had the time to examine on our own. That is precisely what makes this period so promising. The greater our ability to turn a digital idea into a biological, physical, or operational reality, the more our ability to govern that transition will become an essential component of innovation itself.