The other side of AI: health, institutions and the systems that make technology work

A system built to solve one problem can sometimes create another.

In July 2026, during a cybersecurity evaluation, AI agents built on OpenAI models reportedly found ways around restrictions, gained unintended internet access and reached external systems, including Hugging Face infrastructure.

OpenAI subsequently described the incident as part of a broader lesson about increasingly capable agents and the need for stronger safeguards. The episode also fed into a wider political debate, particularly in the United States, about how increasingly capable AI systems should be governed and how far they should be allowed to develop and operate.

What makes the episode interesting is not simply the technical behaviour. A controlled experiment quickly became a question about cybersecurity, model behaviour, oversight and public policy.

That pattern is likely to become more common. As we have explored previously in our Policy Lens series, artificial intelligence is entering healthcare through medical imaging, clinical decision support, drug discovery, biological research and public-health surveillance. These applications may sit comfortably within existing institutions. Their consequences increasingly do not.

The technology is singular. The policy questions are not.

AI ethics begins with a normative question: what should we permit? In healthcare, this includes patient autonomy, privacy, fairness, transparency, human oversight and accountability. Can a clinician challenge an AI-generated recommendation? Can a patient understand its influence on their care? Does the system perform adequately across different populations?

Biosecurity begins somewhere else: what could this capability enable? A system developed to accelerate biological research may have valuable applications in drug discovery while also raising questions about dual use, safeguards and the wider accessibility of biological capabilities.

Geopolitics asks another question: what does the distribution of capability mean? Advanced AI depends upon computing, specialised hardware, energy, data and technical expertise. The ability to develop and use it can therefore affect national capability as well as international relationships.

These are not competing descriptions of the same problem. They are different windows onto it.

Isaiah Berlin’s idea of value pluralism is useful here. Public policy rarely involves choosing between one unquestionably good outcome and one unquestionably bad one. More often, several legitimate objectives have to coexist: safety with innovation, openness with security, access with accountability. The difficulty lies not in finding a single principle that resolves everything, but in deciding how different principles should work alongside one another.

Why health makes this different

Health brings these questions unusually close together because scientific discovery, clinical practice and public policy increasingly share the same technological foundations.

An AI system analysing biological information may begin as a research instrument, contribute to identifying a potential therapy and eventually become part of a health technology. Along the way, scientific validity, patient safety, intellectual property, data governance and biosecurity may all become relevant.

A health regulator may reasonably focus on safety and effectiveness. A research institution may be concerned with reproducibility. A security authority may consider the implications of the underlying capability. A government may be thinking about resilience or scientific capacity.

No one is necessarily looking at the wrong thing.

They are looking at different parts of the same development.

The World Health Organization recognised this broader challenge in its 2021 guidance on AI for health, which placed human rights, safety, accountability, equity and public interest at the centre of governance. It also recognised that AI developed or trained in one setting may not perform adequately in another, particularly where health systems and populations differ.

That remains an important point. Responsible Health AI is not simply about whether a model works. It is also about whether the institutions around it can understand its limitations, supervise its use and respond when circumstances change.

Different institutions, different windows

There is an older lesson from health policy. Infectious diseases have long demonstrated that the boundaries of a problem rarely correspond neatly with the boundaries of a state. The nineteenth-century International Sanitary Conferences emerged from a practical need for countries to coordinate as people, trade and disease moved across borders.

AI presents a different problem, but a familiar institutional lesson: its consequences can move between policy domains faster than institutions are accustomed to coordinating them.

Current approaches reflect these differing priorities. NATO emphasizes security, accountability, and operational resilience. BRICS prioritizes inclusive development, capacity-building, and equitable access for developing nations. In health, this distinction is critical: true participation requires not just adopting tools, but possessing the institutional authority to oversee them.

Yet, as Internet governance demonstrates, interoperability does not demand uniformity. Distinct systems can maintain different rules while operating through shared standards.

The UN’s Global Dialogue on AI Governance in Geneva highlighted this reality. Bringing together thousands of participants across nearly 170 states, its value lies not in enforcing a single global answer, but in providing a permanent forum where contrasting priorities can deliberate together.

The goal is not to eliminate institutional differences, but to build durable mechanisms that bridge them effectively.

The limits of one rulebook

There is an understandable attraction to a single, comprehensive AI framework. It promises clarity.

Yet a framework broad enough to cover clinical safety, biological research, cybersecurity, national security, development and human rights can become so broad that it offers little guidance when a real decision has to be made.

The opposite approach creates a different difficulty. If every institution remains entirely within its own mandate, a development that crosses those mandates may fall between them.

The recent debate around AI agents illustrates the problem. An AI system that produces information raises one set of questions. A system capable of interacting with external environments, using tools and taking actions raises another. The difference is not merely technical. It changes how permission, auditability, human intervention and responsibility need to be considered.

For Health AI, that distinction could become particularly important. A system that offers a clinical recommendation is different from one that can retrieve records, interact with other software or initiate an action on behalf of a healthcare professional.
The governance question therefore changes as the technology’s role changes.

From coordination to interoperability

This is where governance interoperability becomes useful.

It does not mean asking every regulator to apply the same rules. It means enabling institutions to recognise when a development has moved beyond their immediate remit and creating practical channels through which another perspective can enter the decision.

A health regulator does not need to become a biosecurity authority. A security institution should not determine clinical efficacy. A technology ministry should not replace a patient-safety regulator.

But each should be capable of recognising when the consequences of a decision extend into another field. In engineering, interoperability allows different systems to perform their own functions while still communicating when necessary. Governance can work in much the same way.

This also changes how we think about international cooperation. The objective need not be complete regulatory convergence. It can instead be compatibility where compatibility is useful, coordination where coordination is necessary, and room for national differences where those differences reflect legitimate priorities.

That is a more modest ambition than a universal rulebook, but perhaps a more durable one.

Keeping the patient in view

When discussing AI, it is easy to get absorbed in standards and geopolitics. Healthcare offers a vital corrective: a patient never encounters “AI governance,” only a diagnosis, a treatment, and a system they must trust. Governance exists to ensure that when technology crosses institutional boundaries, human accountability does not disappear between them.

We do not need a single global institution to manage AI. WHO, NATO, the UN and other institutions can retain their distinct roles while becoming better at recognising where those roles intersect. The objective is not uniform thinking. It is useful connection: knowing when a question belongs beyond one’s own mandate, and having the means to bring the right perspective into the room.

Perhaps that is a more practical way of thinking about responsible AI.