Podcast 🎧 & blog: National Cyber Resilience in the Age of AI

Written by Federico Plantera, Researcher and Journalist

In March 2026, the cybersecurity landscape reached a turning point when Anthropic and OpenAI unveiled AI models so powerful they could penetrate weak spots in cyber defences, and the AI labs released them only to trusted firms. This exponential take-off has since redefined the digital battlefield, shattering the assumption of linear AI progression.  

For Madis Tapupere, Former CTO of the Estonian Government, now a Tech Management Consultant, that moment set him and his colleagues from government, industry, and academia to work on a question: if offensive capability was about to become radically cheaper and more widely available, what should a state do about it?  

Their work resulted in National Cyber Resilience in the Age of AI: A Doctrine for Asymmetric Cyber Defence, the white paper we are exploring in this episode hosted by Merle Maigre, Head of Cybersecurity at the e-Governance Academy (eGA). 

What is genuinely new in the white paper? “A lot of stuff that has been kind of preached so far has just become more relevant,” Tapupere admits. “We need to do a lot of things we have been talking about for years – just much better, now.”  

What the state cannot afford to lose 

The paper opens with a simple but alarming trend. Once a software vulnerability becomes public knowledge, how long before someone actually exploits it? In 2018, the answer was about two years, on average. By 2023, it was down to five days. By 2026, attackers sometimes exploit a flaw before it is even publicly disclosed. What used to take a skilled team months of dedicated work can now be done in hours, by far more people than before: not just the handful of state intelligence agencies that once had the resources for serious cyberattacks. Against that backdrop, the paper sets out six policy moves a state should make.  

The first two are about what to protect, and how. A government that tries to defend everything equally ends up defending nothing well. So, the first move is to draw a line: designate a small set of functions – identity, financial settlement, emergency services, constitutional continuity – that the country cannot afford to lose for even an hour and hold only those to the very highest standard. The paper calls this the Minimum Viable State. Much of that infrastructure, though, sits in private hands, not public ones.  

Pressed by Maigre on the resulting tension, Tapupere treats it as familiar territory: Estonia has worked through exactly this question since its data-embassy initiatives and during his own tenure as CTO. “It’s never a joyful thing for a private sector company to be loaded with additional regulatory pressure,” he concedes, “but it seems to be working as a system.” 

The second move follows directly from the first: a strong country needs the National Trust Backbone. That is a connected network of reliable public and private services, such as identity, signature, registers, exchange, audit, hosting. It that way every official government action is fully trusted and legal. Naming a small set of critical functions, though, also creates an obvious target: compromise the system everyone depends on, and the payoff is enormous. “You don’t avoid it,” Tapupere says plainly. “You can just address the protection better.”  

His answer is redundancy, not retreat: Estonia’s architecture is already distributed rather than piled into a single data centre, and the doctrine formalises that discipline. The paper’s own illustration is the 2017 flaw found in the chip used on roughly 750,000 Estonian ID cards, which forced a precautionary mass suspension of certificates. The country kept functioning regardless, because it had never relied on a single means of proving who someone was. 

The necessity to keep up pace with machine-speed attacks 

The next moves in the doctrine address the threat itself: what AI actually changes about attacks and how quickly defenders now need to move in response. “AI is a tool – it speeds up vulnerability detection, but it also provides automation to the attackers.” Capability that was once reserved to a handful of nation-state teams has become “very democratised, available to almost anyone, not just sophisticated adversaries,” Tapupere says. From his own experience running internet banking security, he once traced a wave of fraud to two twelve-year-olds struggling at school, armed with nothing more than scripts downloaded from the internet – a preview, he suggests, of exactly this democratisation. 

Maigre’s own reference lands the same point from a different angle: British law enforcement, she recalls, now jokes that APT – once shorthand for Advanced Persistent Threat (a sustained, well-resourced intrusion associated with nation-state intelligence services) – increasingly stands for Advanced Persistent Teenager. The relabelling is tongue-in-cheek, but the underlying shift is not. What required a state-backed team and months of preparation a decade ago can now be carried out by teenagers with downloaded scripts – and that was before generative AI made the toolbox cheaper still. 

If attacks have sped up, the paper argues, defence has to speed up to match. AI side of the paper rests on the idea that while the attackers are using AI to reduce the costs and scale malicious operations, the economics of cyber conflict should be changed in favour of defenders. Fixing software flaws once a month, or even once a week, was good enough when attackers needed months to exploit them. It is not good enough when a flaw can be weaponised within hours. “This ability to make sense fast, react fast, decide fast, and patch fast needs to be on the defender side,” Tapupere argues. Challenged by Maigre, asking how AI-enabled defence can reverse this imbalance if the attackers can always use AI without worrying about regulation, accountability, false positives or disrupting essential services, Tapupere explained that things can productively contribute to automate defence:  

• Legal authority for defenders to act inside national jurisdiction without waiting for case-by-case approval,  

• A constantly updated map of everything the state runs online,  

• AI-assisted tools, governed as carefully as any other critical system, to help spot and fix problems faster than a human team could alone. 

First? Telling the truth 

But another argument runs beneath the technical one, and that’s what Tapupere returns to most. Hostile campaigns against a democracy rarely aim to destroy infrastructure outright. Their actual target, the paper argues, is whether citizens still believe their government functions – a lineage Estonia knows well, tracing back to the 2007 attacks that first made the country a frontline case. The white paper puts the stakes sharply: such a campaign “is fought to make a country look broken to its own citizens, and the government slower than the news cycle in answering for it.” 

In practice, this means disclosure before containment: incident templates written and rehearsed ahead of a crisis, so the state’s own account reaches citizens before a rumour can spread. The paper points to Sweden’s Psychological Defence Agency, and Taiwan’s norm of a signed government rebuttal within roughly an hour of a fabricated claim.  

“Should Estonia’s cybersecurity authority start measuring itself by public confidence, rather than recovery times alone?” Maigre asks. Tapupere is sceptical that trust can be captured in a single metric.  

Every person working in the cybersecurity authority should see themselves as responsible for maintaining public trust, he argues, but the authority cannot do this alone. In Estonia, population registries sit under the Ministry of Internal Affairs, business registries under other ministries, each responsible for protecting its own data. And private sector too, though, holds sensitive data.  

The conclusion is that cyber resilience is not one agency’s mandate but a shared discipline running through every register, every ministry, and every company handling citizens’ data. “Whenever our digital government fails in some corner,” Tapupere says, “the trust starts eroding.” 

Looking out five years from now, our guest sees AI’s trajectory as echoing the early internet: built fast, and secured only once the danger became undeniable. “I think we need to run very fast to stand still,” Tapupere concludes. “If things won’t have gone significantly worse, then we are already in a very good place.” Not the most optimistic of views, perhaps, but definitely a serious and compelling work plan for the years ahead. 

Interested in more?