Google DeepMind has launched a bioresilience programme, an initiative aimed at deploying artificial intelligence to defend against biological threats. Announced together with Isomorphic Labs, its sister company focused on drug discovery, the project has a stated dual objective: prevent malicious use of its models, and at the same time offer governments, scientists and biosecurity organizations AI tools to anticipate and neutralize pandemic risks and bioterrorism.

The move did not come out of nowhere. With AlphaFold, DeepMind has already shown how AI can accelerate the understanding of proteins—a fundamental building block also for designing new molecules, including potentially harmful ones. Isomorphic Labs, for its part, is pushing AI into drug discovery. Together, the two entities possess expertise that, while promising to revolutionize medicine, also inevitably attracts those seeking to develop biological weapons. Bioresilience, then, is not just a research programme: it is an acknowledgement of the dual-use nature inherent in these technologies and an attempt to build a safeguard before it is too late.

But beyond the noble intent, the initiative shines a spotlight on a crucial issue for the entire AI ecosystem: who physically controls models and data when national security interests are at stake? In the biosecurity field, datasets are hypersensitive: genetic sequences, targeted protein structures, simulations of artificial pathogens. Entrusting them to external cloud services, even from a giant like Google, means accepting a transfer of information sovereignty that many states and government agencies are unwilling to concede. It is no coincidence that the programme explicitly mentions support for governments and biosecurity groups: the real use of these technologies will almost certainly require on-premise deployment, in air-gapped environments or on tightly controlled infrastructure, far from commercial data centers.

The issue of data sovereignty is not abstract. For an agency simulating a response to an outbreak, every piece of information that travels over shared networks represents a risk vector. And the models themselves, if run as a cloud service, could expose usage patterns through their APIs, revealing sensitive intentions. DeepMind, aware of these criticalities, may be pushed to offer containerized versions of its tools, suitable to run on local hardware, rather than keeping everything behind a proprietary endpoint. This is a dynamic we have already seen in finance and healthcare, and now it extends to biosecurity.

Structurally, the programme marks another step in the race toward proactive AI regulation. We are no longer in the "move fast and break things" era: leading companies are building self-imposed safety fences, partly to preempt government regulations, partly to manage reputational risk. But this creates a paradox: the more access to the most powerful models is filtered through safety mechanisms, the more decision-making power is concentrated in the hands of those who design those mechanisms. Who decides what constitutes "malicious" use? An academic researcher exploring bioterrorism scenarios to understand defense strategies could be blocked by the same filter meant to stop a bad actor. The line between defense and offense, in biosecurity as in AI, is thin.

For those today evaluating deployment architectures for AI workloads involving sensitive biological data, the choice is no longer purely technical, but one of responsibility: owning the entire stack, from GPU to model, becomes a non-negotiable requirement to ensure that no data leaves the control perimeter. AI-RADAR tracks these evolutions, offering analytical tools to navigate the trade-offs between cloud and on-premise, aware that security in these domains cannot be delegated.

And meanwhile, DeepMind's message is clear: AI can be a formidable ally against biological threats, but only if the international community can build trusted, distributed infrastructures capable of operating under the full control of those responsible for protecting public health.