The AI cyber threat is here
Warnings of autonomous agents taking over the internet may distract from a more immediate danger: governments and criminal groups are already using the technology to attack critical infrastructure.
The year is 2027. You wake up one morning to find you cannot use email or send instant messages. Your online bank account is down too. You turn on the TV to find a world in meltdown.
Hundreds of organisations worldwide have been hacked, locking thousands of employees out of critical systems. Eventually, public transport breaks down and industrial plants explode. Governments hold emergency meetings to establish who is responsible. This time the attacker is not an enemy or a terrorist cell, but a thing. The AI agents are in charge now.
This, roughly, was the vision of the future conjured up two months ago by Dario Amodei, chief executive of Anthropic, when he warned in a 4,000word essay that a swarm of agents could be capable of taking over the entire internet within six to 12 months.
For some, parts of the prophecy already seem to be coming true. In the past week, IT systems at organisations including Microsoft, JPMorgan, the US Federal Aviation Administration and even the US National Security Agency have all crashed, adding to a sense of AI induced chaos. Meanwhile, the cyber arm of the UK intelligence services has been securing tens of thousands of British fridges and freezers.
The unusual series of outages and alerts has arrived at a moment when the number of cyber attacks is already at record highs. So far this year, hackers have been accused of crimes as varied as the assaults on European supermarket freezers that sparked the UK into action, compromising control systems in American water utilities, breaking into the FBI and stealing $6.3bn from the crypto exchange Binance in what was the largest robbery in history.
For cyber security experts, though, the threat now posed by AI is not selfinflicted network outages or Amodei’s prediction of an all-powerful autonomous agent — or indeed an extinction level event for humanity.
The more pressing concern is that states and criminal groups are already using the technology to support and enhance offensive capabilities at a pace and scale the world is not prepared for.
“It’s not as bad as people want to imagine, but it is worse than most people realise,” says Robert Lee, chief executive of cyber security firm Dragos and a leading authority on cyber attacks.
Lee is one of many cyber professionals who say the focus by AI leaders on theoretical worst-case scenarios is distracting people from the very real danger that AI-assisted cyber crime and cyber warfare already pose.
“Is defence doable? Yes,” he says. “Can we win? Yes. Are we going to? No.”
Extortion, phishing and self-inflicted wounds
The frequency of cyber attacks was already rising fast even before the AI revolution. Charl van der Walt, global head of security research at Orange Cyberdefense, part of the French telecoms group, says the cyber extortion incidents tracked by his team — attacks where hackers steal data for sale or ransom — have multiplied over the past six years from roughly 1,500 in 2020 to 7,100 in 2025. Over that time, he adds, the number of active cyber extortion groups has risen from 33 to 114.
The rapid growth has been driven by factors including increasing digitalisation, the professionalisation of cyber crime and the rise of cryptocurrencies that have made it easier for hackers to get paid. This year, van der Walt has already tracked 7,500 such hacks in the nine months to September. While some of those attacks may have used AI tools in some way, none were automated hacks run by AI, according to Orange’s research. “AI technology is not driving the security problem; it is revealing the security problems,” he says.
In August, American researchers used OpenAI and Google models to identify vulnerabilities in PHP, a programming language behind 70 per cent of all websites, that had been undetected for more than a decade.
Organisations are now identifying software bugs at such a rate that IT systems cannot keep up with the pace of the necessary fixes.
US officials have said the rush to update software was behind the recent outages that blocked access to Microsoft email in 70 countries, grounded all air traffic in the country and even disrupted a classified internal network at the NSA, the intelligence agency that monitors and intercepts electronic communications. “Every day brings another reminder that AI is moving faster than our ability to manage its consequences,” Mark Warner, vice-chair of the Senate Select Committee on Intelligence, said of the recent outages.
US President Donald Trump was more sanguine. “Tell them to relax, everything’s good,” he told Fox News on Wednesday as he sought to assure Americans that the US was not under attack. “They’re putting in new equipment, and things happen when you put in new equipment.”
Such assurances did little to quell conspiracy theories that the outages were caused by co-ordinated cyberterrorism — or concern about the vulnerability of vital infrastructure. By Thursday, Trump had sacked US national cyber director Sean Cairncross.
For now, says Marcus Hutchins, a British cyber security researcher who has first-hand experience of cyber crime, hackers are mostly using AI tools such as ChatGPT to develop and execute phishing attacks through which people are tricked into handing over their login credentials.
The exception, Hutchins acknowledges, is “top tier” state threat actors, such as China, Russia and the US, who he suspects are “stockpiling” the most advanced zero-day exploits for future deployment: “A lot of their work is around building capabilities for the future, so we don’t really see what they have, only what they use.”
Matters of life and death
The biggest shift so far in the balance of power between hackers and defenders, experts say, concerns operational technology, known as OT.
Most hacks target information technology systems with the aim of stealing money or data. OT attacks target industrial equipment such as power plants and factories with the objective of locking out operators to demand a ransom or cause damage.
Historically, such attacks have required highly specialised knowledge and required months of expensive planning. Perhaps the most famous OT attack of all time, the Stuxnet malware assault on Iran’s nuclear programme by the US and Israel in 2009, took years of preparation and involved building a physical replica of Iran’s centrifuge systems to test the malware before deploying it.
Hacking groups can now use large language models to read the manuals of industrial equipment and conduct research on past accidents, says Lee at Dragos. OT is often poorly defended, precisely because it was technically complicated to attack, he notes.
“When Dario [Amodei] comes out and says, ‘Hey, I think people are going to die’, I agree with him, but for different reasons,” Lee adds. The risk will come from human adversaries using AI models to sharpen their attacks on critical infrastructure, he says. “That intersection is dangerous.”
Lee notes that sophisticated offensive cyber capabilities have moved from intelligence agencies in the 2010s to the military in the 2020s to a motley mix of proxies, government contractors and criminal groups today. Where intelligence agencies tend to use such tools sparingly, preferring to conduct deniable, covert operations, the newest actors lack the same restraint. “I think it is a very safe assessment to say we are going to see loss of human life as a result of OT cyber attacks,” he says.
When the front line is a freezer
Many western countries are already in conflict, either officially or unofficially, with major cyber powers. Since the US and Israel began air strikes on Iran in February, Iranian hackers have staged offensive cyber operations on US domestic targets, including attacks on dozens of water companies. Last week, a cyber unit of Iran’s Islamic Revolutionary Guard Corps was blamed for October’s unprecedented cyber attacks in France, which turned up the temperature on hundreds of supermarket freezers to as high as 50C, destroying millions of euros of food. Similar attacks hit supermarkets in 10 other countries across Europe this month, prompting the NCSC’s discreet nationwide effort in recent weeks to safeguard British refrigeration.
“I don’t think there’s going to be some kind of conscious use by a state of largescale AI attacks outside of conflict, as the rules of engagement still apply,” says a western intelligence official says. “Where you have greater risk is that all our understanding of the capability is still nascent. Everybody’s still learning, so there is more likelihood of unintended consequences.”
European officials are increasingly open about Russia’s use of so-called hybrid warfare, including cyber attacks, against EU member states to exert pressure related to its war against Ukraine. Iran and Russia’s strategy is to “undermine the confidence and the cohesion of the society that you’re in conflict with”, says Orange’s van der Walt.
China has a different approach. Western officials say Beijing is more interested in capabilities that enable espionage or serve as a deterrent against foreign attacks on its own people and infrastructure.
In a far-reaching cyber campaign exposed in 2023, US authorities found Chinese malware embedded in American energy, water, pipeline and transportation systems in apparent preparation to disrupt, if needed, the US economy and military operations. The following year, US agencies made an even more shocking admission. Chinese hackers had gained and maintained access to US telecommunications networks for several years, giving the country’s intelligence potential access to the unencrypted phone communications of almost anyone in America.
For Anne Neuberger, a former US deputy national security adviser who once led intelligence operations for the NSA, China’s control of its internet and economy gives it a natural advantage in cyber warfare:. “The tension between an authoritarian country’s capabilities in terms of being able to monitor and defend and a democracy’s [ability to do so] become[s] even sharper in the age of AI,” she says, noting that the Chinese government has a direct role in protecting critical infrastructure.
She is less concerned about the vulnerability of big western banks with large cyber security budgets than about infrastructure such as hospitals, pipelines and power systems, which were historically expensive to secure and have never been properly protected.
These are, she adds, “systems that were built over the last 50 years, well before sophisticated offensive cyber capabilities”.
Cyber security companies are already all rolling out AI-powered defence technology. Some help security teams to identify and analyse threats. Others use agentic AI systems to respond to and neutralise attacks. Another possibility, Neuberger says, is to build digital twins of real-world critical infrastructure on which the latest frontier models can be tested. As the cyber wars escalate, she maintains there is no time to waste in using the new technology to protect, as well as attack, the infrastructure that has become the new battleground: “AI absolutely helps attackers . . . but it’s a game-changer on defence.”
Comments
Post a Comment