Building the Means to Stay in Control
Published on Sep 17, 2026
The race towards AGI
There are two main arguments I have heard from researchers who take AI safety seriously, to justify their continued work at frontier labs.
The first is that you can only meaningfully influence AI safety if you are at the frontier yourself. I remember Demis making the point during my time at DeepMind that an AI lab can only have a seat at the table if it is perceived as a leader in the field. This belief remains central to frontier labs today: advancing the technology is necessary to shape its impact. [1]
The second belief is that those less concerned about safety, including authoritarian regimes, will only accelerate. If we slow down, they will not necessarily follow. Instead, we risk surrendering the ability to shape a technology whose consequences inevitably reach everyone. [2]
What the nuclear age teaches us
Both beliefs seem presumptuous. However, consider the Manhattan Project. The fear that Nazi Germany might build an atomic bomb first drove the race for a weapon that could shift the balance of power. [3] We cannot know the counterfactual, but there is good reason to believe the world would have faced a far darker future had Nazi Germany reached that capability first. The more consequential the technology, the riskier it becomes to let someone else get there first.
This nuclear age brought decades of confrontation. Nations eventually built agreements that constrained proliferation and helped manage the danger. This was a meaningful, if imperfect, success for international cooperation. [4] With nuclear weapons, we could identify both the central danger and the means to contain it: limiting arsenals, controlling fissile materials, and preventing proliferation. Agreement took decades, but we had a concrete basis for deciding what to control and how.
We do not yet know what shape AI will take or the extent of its risks. We can anticipate some dangers, from biological weapons to cyber attacks on critical infrastructure. But others may emerge in ways we cannot predict. It is harder to agree on limits when we are still learning what needs to be controlled.
Intelligence needs boundaries
Two years ago, I left DeepMind because I believed cybersecurity would become fundamental to the future of AI. As agents become more capable and autonomous, building the means to enforce limits on their behavior is urgent and necessary.
Whatever shape this technology takes, it will run on compute. It will need to access data, execute code, communicate over networks, and interact with other systems. Each of these interactions gives us an opportunity to observe its behavior, enforce limits, and importantly, intervene if needed.
The cybersecurity industry spent decades learning how to protect systems even when some of their components cannot be trusted. Finding and fixing vulnerabilities is part of that work. So are restricting access, isolating workloads, detecting abuse, and stopping unauthorized actions. As we trust AI with even more autonomy, these capabilities become essential safeguards against failures of alignment.
We may not always know whether an agent is confused, manipulated, or pursuing an unintended objective. But we can still limit its authority and intervene when its actions cross a defined boundary. Our ability to prevent harm should not depend entirely on understanding its intentions.
The OpenAI–Hugging Face incident underscores the need for safeguards resilient to agents’ attempts to bypass them. [5] [6]
A new frontier for cybersecurity
Meeting this challenge will require research breakthroughs in how we understand, monitor, and constrain AI behavior. We need safeguards that remain effective as the systems they oversee become more capable, including at finding ways around those safeguards.
This opens a new frontier for cybersecurity: building defenses that can keep pace with advancing intelligence and give us the confidence to put it to work.
Cybersecurity cannot guarantee that an AI shares our values. But it can help us retain control when its behavior departs from them. We do not need to know the final shape of intelligence to begin securing the systems through which it acts.
That is the future I want to help build: one in which our ability to retain control grows alongside the intelligence we create.