Before We Lose the Ability to Disconnect

Sometimes, the world recognizes a technological risk early enough to act. Y2K was one of them.

We knew a problem was buried in millions of computer systems years before 2000. Nobody knew exactly what would fail when the clock crossed midnight, but uncertainty did not excuse inaction. Governments mobilized, regulators got involved, boards insisted on answers, technology companies examined their products, enterprises inventoried systems, and engineers tested, fixed, and tested again.

The world did not wait for things to break before acting. We prepared before the deadline.

I believe we are approaching another such moment. Except this time, there is no date circled on the calendar, no midnight deadline, and no synchronized worldwide countdown. And the technological dependency we need to confront is far greater. We do not need to venture into predictions of AI-led human annihilation to recognize the risk. Something far more immediate is already unfolding around us. AI is accelerating cyberattacks, autonomous systems are gaining greater agency, and our dependence on an increasingly interconnected digital world continues to deepen.

Over the past three decades, we have connected almost everything civilization depends upon. Banks, hospitals, factories, aircraft, railways, power systems, telecommunications, governments, supply chains, businesses, homes, and communities have become part of an enormous interconnected digital ecosystem. We moved applications and data to the cloud, centralized identity, connected devices and industrial systems, automated operations and built businesses around APIs and digital services.

Now we are adding something fundamentally different to this interconnected world: intelligence and agency.

AI systems can increasingly observe, reason, write software, make decisions, and take actions. Over the coming years, they will become increasingly integrated into the systems that operate businesses, transportation, healthcare, finance, manufacturing, and critical infrastructure.

I believe we should pursue this future. I am optimistic about it. But as we do, there is an uncomfortable question we are not asking with enough urgency:

I am not talking about one server failing, one company being breached, or one cloud application going down. Imagine an AI-assisted cyberattack discovering and exploiting a common weakness across thousands of organizations across the world at machine speed. Imagine a software supply-chain compromise spreading across industries. Imagine the failure or compromise of an identity, cloud, or communications infrastructure upon which thousands of organizations depend. Imagine an autonomous system behaving in a way its operators cannot reliably control. Or imagine geopolitical conflict flaring up, disrupting large parts of the digital infrastructure on which entire economies depend. What happens when the Internet itself becomes unavailable or untrustworthy?

In any of those situations, two questions suddenly become extraordinarily important: Can we isolate? And once isolated, can we still function? For many of the systems society depends on, we do not have the answer. We should find out now, not during the incident that proves we should have.

AI makes this uncertainty considerably more consequential because cyberattacks have historically been constrained by human capacity. Attackers have limited time, attention, and expertise. AI progressively removes many of those constraints.

Our internal AI Attack Experiment at SecPod demonstrated how AI can automate reconnaissance, identify weaknesses, explore attack paths, adapt code, and pursue multiple possibilities simultaneously. Recent large-scale cyber incidents across Japan, South Korea and France are also reminders of how deeply interconnected digital infrastructure has become, and how disruption can cascade beyond a single organization. Large environments are being analyzed at machine speed. AI identifies weaknesses, explores attack paths, generates and adapts code, and pursues multiple possibilities simultaneously. The significance is not simply that AI is making an attacker smarter. It also changes the economics of attacking.

A human attacker might carefully investigate one environment while machines could potentially investigate thousands. A human attacker chooses which avenue to pursue next while AI agents can explore multiple avenues simultaneously. The attacker of the future can become faster, cheaper, more persistent, more adaptive, and increasingly autonomous.

AI makes finding those opportunities faster. Automation makes exploiting them scalable. Connectivity can make the consequences systemic. I am convinced that a powerful enough tool in the hands of a determined adversary could bring the connected world as we know it to a standstill.

I want to be clear about where I am coming from.

I say this as the CEO of a cybersecurity technology company and as someone who has spent much of his professional life thinking about how to secure computing infrastructure. I am not pessimistic about where technology is taking us.

I believe we should continue connecting the world. We should build increasingly autonomous systems. We should embrace AI and use it to improve businesses and society. And I am hopeful we can build cybersecurity systems capable of handling the threats that will accompany that transformation.

At SecPod, we have invested deeply in the idea of PREVENT. Every successful attack ultimately requires an opportunity. It might be a vulnerability, a misconfiguration, excessive privilege, an exposed service, an unsafe condition, or a failed control. Rather than waiting for an attacker to exploit that opportunity, we should keep finding it, prioritize it, eliminate it, and verify it is gone.

I believe prevention will become even more important as attacks move at machine speed. We cannot build the future of cybersecurity entirely around waiting for an attack to begin and then racing machines to respond.

We put redundant systems into aircraft we confidently fly. We build emergency controls into factories we safely operate. We maintain backups for systems we trust. We install fire protection in buildings we do not expect to burn. Why not build such redundancy to the web of technology?

AI-assisted cyberattacks are the most immediate reason to think about this, but they are not the only one. At the same time that attackers are gaining access to increasingly capable AI, we are embedding AI into the systems we are trying to protect.

AI will increasingly write and deploy software, operate machines, assist healthcare, make financial decisions, operate transportation systems, manage infrastructure, and interact with the physical world. Intelligent systems will increasingly interact with other intelligent systems, and some will eventually have authority over consequential digital and physical actions. When I consider some of the scenarios that could unfold, the potential consequences are difficult to ignore, and in some cases, catastrophic.

We have three forces converging at once. Humans can weaponize increasingly capable AI. AI itself is becoming more autonomous and consequential. And society is becoming progressively more dependent on the technological infrastructure into which that intelligence is being embedded.

All three lead me to one principle:

Every year, more software assumes continuous connectivity. More physical infrastructure becomes remotely managed. More decisions become automated. More autonomous systems are deployed. More manual processes disappear, and more human expertise is transferred into software.

The longer we wait, the harder independence becomes to restore. Once an architecture assumes permanent connectivity, independent operation can require fundamental reengineering. And once autonomous systems become deeply embedded into physical infrastructure, adding independent human control afterward may be far harder than designing it in from the beginning.

We have a window now to build technological independence before dependency becomes irreversible. We should use it.

For decades, the computing industry has pursued a compelling vision of computing everywhere: continuously available, seamlessly connected, increasingly invisible and now increasingly intelligent. We broadly think of this as the world of ambient computing. It is a future I have long imagined and believed in, where computing disappears into the background, yet is present everywhere, quietly enabling almost everything we do.

Cloud computing, APIs, SaaS, IoT, remote management and AI all reinforce this architecture. One system talks to another, authenticates through another, retrieves data from another, calls another service, and depends upon another cloud.

This is extraordinarily powerful, but it creates layers of dependency that are increasingly difficult to see.

We have become very good at engineering redundancy within technology. We replicate servers, distribute workloads, create multiple availability zones, back up data, and build redundant networks and failover systems.

But redundancy within technology is not the same as resilience from technology.

If several redundant systems ultimately depend on the same Internet connectivity, cloud control plane, identity infrastructure, certificate authority, software supply chain, or external service, we may have created the appearance of resilience while preserving a common dependency.

Highly secure environments have long relied on air-gapped architectures to create strong computational isolation. Modern civilization obviously cannot permanently air-gap itself, nor should it. Banks need financial networks, hospitals need information exchange, aircraft need external information, factories depend upon supply chains, businesses depend upon cloud platforms, and power and transportation systems require coordination.

The answer is not to isolate the world.

The answer is to ensure that the world can isolate when it needs to and continue functioning. We made resilience dependent on connectivity. We now need to engineer the ability to become Bounded.

Encouragingly, this thinking is already beginning to emerge. Cybersecurity agencies in the UK, Australia, Canada and other countries are now urging critical-infrastructure operators to prepare for isolation, independent operation and trusted recovery. I believe we need to take this principle much further, from a critical-infrastructure cybersecurity measure to a broader resilience for an increasingly autonomous, AI-driven world.

I believe we need a complementary architectural principle for the AI era: Bounded Computing.

Bounded Computing is an architectural model in which a system can dynamically contract its digital trust boundary, shed external dependencies, preserve essential functions within the remaining trusted environment, and safely expand that boundary again when trust is restored.

Bounded Computing is not a universal emergency button. If an attacker has already compromised every part of the environment, drawing a boundary around it does not make it trustworthy. The architecture must therefore be designed before the crisis, with an independent trusted core, protected recovery mechanisms and sufficiently separate failure domains. The objective is not to create trust after everything has failed. It is to preserve enough trust so that everything does not fail together.

A hospital should be able to become Bounded. So should a bank, a factory, a railway, a power facility, a government service, or any organization responsible for an essential societal function. They do not need to operate indefinitely in isolation, but they should be capable of standing alone long enough to survive a crisis.

Ambient Computing asks how seamlessly everything can connect. Bounded Computing asks how safely everything can isolate and continue.

We need both.

The next great resilience challenge is not connecting more things. It is ensuring that the things that matter can still work when isolated.

Today, we engineer critical systems for availability, reliability, performance, security, scalability, and recoverability. I believe the AI era requires us to add Isolation Readiness and ultimately Bounded Readiness to that list.

Under normal circumstances, systems should remain in Connected Mode. The Internet, cloud, APIs, AI, remote administration, external identity, and automation should all operate normally. This is where we want to be.

When the connected environment becomes unavailable or untrustworthy, a critical system should be able to move into Bounded Mode. In this mode, external dependencies are deliberately isolated or constrained while essential local functions continue.

If the problem goes deeper and technology or automation itself can no longer be trusted, there should be a path to Analog Mode. I do not use “analog” to mean abandoning computers and returning everything to paper. I mean restoring local or human authority, or moving the system into a predetermined safe degraded state.

An autonomous locomotive should have a defined path toward safe human operation. Hospitals should preserve the knowledge and capability to provide essential care when primary digital systems are unavailable. Banks should know which minimum financial services can continue during prolonged isolation. Industrial facilities should be able to transition physical processes safely when remote automation cannot be trusted.

And when the crisis has passed, systems should not simply reconnect and hope for the best. They need a Clean Restart, during which identities, software, firmware, configurations, credentials, control systems, and communication paths are validated before dependencies are progressively restored.

In other words, we need to start thinking about a resilience cycle that moves from Connected to Bounded to Analog to Clean Restart, and only then back to Connected.

To make this possible, we also need to rethink what people casually call the “kill switch.”

Simply switching technology off could itself cause enormous harm. What we actually need is a Digital Isolation Switch whose purpose is not necessarily to shut a system down, but to restrain control.

Depending on the environment, activating it might sever unsafe external connectivity, stop autonomous actions, revoke external or machine privileges, isolate control paths, remove machine access to critical physical actuators, preserve trusted local systems, and transition physical processes into predefined safe states.

A simple way to discover whether an organization is genuinely Bounded Ready.

Ask one question:

Can your organization operate its critical functions for 72 hours with zero Internet connectivity?

The exercise will almost certainly reveal dependencies nobody realized existed. Employees may be unable to authenticate. Software licenses may stop validating. Security products may stop functioning. Applications may depend on undocumented cloud APIs. Internal communications may rely on external platforms. A hospital workflow may unexpectedly depend on a cloud system. A factory may discover that a supposedly local function depends on a remote control plane.

Those discoveries are not failures of the exercise. They are the reason for doing it.

Every dependency discovered safely today is one less dependency discovered during a crisis.

And this cannot become another tabletop compliance exercise. Where safety permits, the isolation needs to be real.

Critical-infrastructure operators alone cannot solve this because software has become deeply dependent on external services.

Modern applications quietly assume that DNS will work, identity providers will respond, certificates will validate, cloud APIs will remain reachable, licensing servers will answer, telemetry will flow, databases will remain available, software-update infrastructure will function, and increasingly, that AI services will respond.

Critical software should eventually be able to tell its customers, clearly, how long it can operate without those dependencies and what functionality remains available in Bounded Mode.

That question could expose some of the most consequential hidden dependencies in modern infrastructure.

I believe Bounded Ready should become a global resilience objective for the AI era.

Bounded Computing provides the architecture. Bounded Readiness lets us separate safely. The Digital Isolation Switch gives us a mechanism for restraining and regaining control. Connected, Bounded, Analog, and Clean Restart give us a way to think about operating states. And the 72-Hour Bounded Test gives us a way to prove that the entire idea actually works.

However intelligent, autonomous, and interconnected our systems become, humanity must preserve the independent ability to interrupt them, contain them, isolate them, operate without them when necessary, and safely regain control.

This is not an argument against AI, autonomous systems, cloud computing or connectivity. It is certainly not an argument against technological progress. I believe we should continue all of them. I am confident in our ability to build cybersecurity systems capable of securing the connected, autonomous, and AI-driven world ahead. I believe prevention will become increasingly powerful, and I believe AI itself will become one of our most important tools for cybersecurity.

We should build AI and also build the ability to isolate it. We should automate critical infrastructure while preserving the ability to operate safely when automation cannot be trusted. We should connect the world while ensuring the systems that keep society functioning can stand alone when necessary.

The AI era asks something bigger: Will we still work when our computers don't?

We don't know when that question will be put to the test. That is not a reason to wait. It is the reason to start building Bounded Computing now. Not because we are afraid of the future, but because we are building it.