There have been dire warnings lately about the threat to humanity from AI. One Anthropic researcher claimed there was a 1-in-10 chance of an AI driven apocalypse in the coming decade, citing the threat from AI with superhuman intelligence [1] with another even claiming the odds may be 50%+ [1a]. So how might this apocalypse play out?
Personally, I am sceptical of the threat of a superhuman AI deciding to kill off or enslave the human race. If an AI system is really so intelligent, it would realise that the data centres and electricity grids that are its nervous system and lifeblood will be dependent on humans, and are likely to be for some time, notwithstanding advances in robotics [1b]. Assuming it has a bias towards self-preservation, it wouldn’t seek to wipe out the human race. My fear is more from mistakes caused by AI systems, and also from AI being mis-used by humans with genocidal intent.
Reverse stress testing where an AI event threatens the human race, I could foresee the following scenarios:
Nuclear War
Thanks to the “Terminator” series of films, the idea of a superhuman AI system triggering a nuclear war is deeply embedded in popular consciousness. However, such a war would be an act of suicide by such a system – the electro-magnetic pulse (EMP) from nuclear strikes would frazzle electricity grids even if some remote data centres weren’t affected by the blasts.
To my mind, a more apt 1980s film analogy would be “War Games” (1983). This stars Matthew Broderick as a hacker who (spoiler alert) breaks into an AI system used by the North American Air Defence Command (NORAD) and starts playing war games with it. This inadvertently leads NORAD to think it is under nuclear attack which almost precipitates a real nuclear war.
Eerily enough, in that same year, Soviet nuclear forces almost started a nuclear war due to a false alarm triggered by satellite warning systems confusing a rare alignment of sunlight on clouds for a missile launch. Only steady nerves by the officer in charge spared mankind oblivion [2].
So, I think there is a real risk that AI embedded in nuclear weapons command and control systems could hallucinate an incoming attack, leading to real life nuclear conflict. For the US, Russia and China, there may be enough time to verify if a supposed launch is real. However, for theatres like India / Pakistan, where capital cities are only minutes from the border as rockets fly, such a hallucination might well trigger a devastating nuclear war killing hundreds of millions.
A lesser threat would be where a deranged individual uses AI to help launch a nuclear war (I am thinking of General Jack D.Ripper in Kubrick’s masterful black comedy “Dr Strangelove” (1964)). AI could for instance be used to impersonate a US Presidents voice to order a nuclear strike. However, I believe other checks and balances should prevent this (though there is the non-AI related risk of a deranged Trump or Putin starting a nuclear war themselves).
AI driven failure of critical infrastructure
Another apocalyptic fear is an AI-enabled cyber-attack causing critical infrastructure to fail. A particular concern is of hackers using AI to find hitherto unknown (or little known) “zero day” vulnerabilities in infrastructure operating software, which they then exploit to bring this down. In the extreme, such an attack might lead to a Chernobyl-style meltdown at a nuclear power plant, or to a sustained outage of the electrical grid, or looking forward, to a shutdown of 5G networks used by self-driving vehicles causing traffic chaos.
I would make five observations about such scenarios. Firstly, cyber-attacks on infrastructure is not a new risk. In 2009, long before the development of Large Language Models (LLMs) like ChatGPT, hackers used a virus called Stuxnet to exploit “zero day” vulnerabilities in Microsoft and other operating systems to cause gas centrifuges used by Iran’s nuclear program to spin out of control [3]. A cyber-attack on nuclear facilities is hardly unprecedented.
A few years later in 2017, the WannaCry ransomware virus caused havoc across a wide range of government and other organisations worldwide [4]. This used a vulnerability in Microsoft Windows operating systems that had been secretly developed by the US National Security Agency (NSA) but which had then been leaked. Amongst other impacts, UK hospitals found themselves locked out of systems, forcing the closure of wards and cancellation of routine operations [4a].
In 2021, there was a cyber-attack on a key fuel pipeline in the US which led to airports and petrol stations running out of fuel, prompting Joe Biden to call an emergency [5]. More recently, there have been numerous attacks on water and other infrastructure in the US which has been linked to the current Gulf War [6].
What AI does is to increase the risk of a successful cyber-attack on infrastructure. In the past year, more sophisticated AI models have shown an alarming ability to identify and exploit hitherto unknown vulnerabilities in systems. In June 2026, the US government placed restrictions on Anthropic’s latest models due to their ability to exploit such vulnerabilities [7]. A few months later, thousands of websites had to be shut down after AI detected vulnerabilities in the PHP language used by these [8]. There is a race between hackers and cyber defenders using AI to identify weaknesses in systems and either exploit or protect against these. In some instances, hackers will get there first to create havoc.
The second observation I would make about cyber-attacks is that while exploiting hitherto unknown vulnerabilities is difficult to defend against, and so attracts a lot of attention, the top vector for most attacks is “phishing”, that is deceiving people into revealing sensitive information or installing malware [9]. Again, AI is exacerbating the threat e.g. through the use of “deepfakes” to impersonate senior management, and/or increasing the scale and sophistication of attacks.
This brings me to my third observation: that whether a cyber-attack is powered by AI or not, the controls are similar. Staff need to be trained in identifying and dealing with phishing attacks. Basic cyber-hygiene should be in place, including keeping software up-to-date and implementing “patches” issued by software providers to guard against attacks.
Unfortunately, this is not always the case. Taking the WannaCry example above, Microsoft was aware of the vulnerability in its systems by the time the attack took place, and had issued a patch to its supported software to protect against this. However, the virus still caused havoc as some organisations failed to implement this patch while others, including a large number of NHS hospitals, were running on unsupported software [4].
A similar situation arose recently when the PHP vulnerability was identified. A patch to fix this was swiftly made available for the latest versions of PHP but many organisations still had to take websites down to update to the latest version of the software.
Considering the recent attacks on US water systems, again the issue is less the sophistication of attacks or the use of AI, and more down to weaknesses of cyber-defences [6]. Hopefully the greater awareness of the threat of cyber-attacks powered by AI will prompt infrastructure operators to invest more in improving cyber-defences.
A fourth observation I would make is that failure of infrastructure could be caused by something other than a cyber-attack. The Chernobyl disaster was caused by a disastrous set of human errors during a safety test [10]. I don’t think it is impossible that AI systems could “hallucinate” and trigger a similar disaster either directly or by misleading human operators into taking a catastrophic course of action.
There is also the risk of system failure. While not AI as such, an example of what could go wrong with AI-directed systems is the recent system malfunction in Wuhan, China, which led to over 100 self-driving taxis coming to an abrupt stop in traffic [11].
My final observation is that while AI-driven infrastructure failure could give rise to mass casualties, it probably won’t cause the extinction of the human race. Disastrous as it was, the most widely cited estimate of deaths resulting from Chernobyl is 4,000 eventual deaths [10] (though some estimates go into hundreds of thousands of deaths). To arrive at an apocalyptic infrastructure failure scenario, you might need to consider whether an AI system could have the same impact of say a EMP or solar flare that might cause a protracted failure in the power grid [12], noting that such a failure would probably lead to the demise of the AI system itself.
Bio-terrorism
The one new risk posed by AI that keeps me up at night is the risk that some madman could use AI to develop a new plague that could decimate mankind. Covid-19 highlighted the death and carnage that a new pathogen can create, but even this wasn’t as bad as the 1918/19 flu pandemic [13] let alone the “Black Death” of the 14th century which killed at least 30% of the population of Europe [13a]. What is particularly worrying about AI is not just its ability to identify new potential pathogens, but also its ability to guide unsophisticated lunatics through the complicated process of creating and disseminating these [14].
There is also the risk that AI enabled research into pathogens gets released into wider circulation. This is an existing risk. There is some suggestion that the Russian flu outbreak of 1977 could have been caused by a laboratory accident [15] while recently there was alarm about the death of a researcher in Siberia from pneumonic plague [16]. There are also some conspiracy theories that Covid-19 was the result of a similar accident at a laboratory in Wuhan.
Looking forward, just as AI promises the bright new future of new treatments for disease, there is also the risk that AI-supported research comes up with lethal pathogens that could be accidently released. For instance, the Economist recently called out the threat posed by “mirror life” biology [17] which is likely to be enabled by AI, and which could be devastating if not properly managed,
Status quo annihilation
Amidst all the hype about whether AI could kill us all, sight has been lost of climate change which is truly an existential threat if not properly managed. After decades in which politicians at least professed to be concerned about climate change, now populist politicians from Donald Trump to Nigel Farage to the AfD in Germany dismiss net zero targets and support continued fossil fuel extraction (“drill baby drill”) and the unfettered growth in energy intensive data centres.
The impact of data centres underpinning AI on energy requirements and carbon emissions is staggering. In their 2026 Global Energy Perspective, McKinsey & Company noted that data centre demand could grow from 800 TWh to between 3,000 – 4,200 TWh [18] by 2030. To put this into context, the lower estimate is roughly equal to one-tenth of current global electricity production of 30,000+ TWh [19].
McKinsey outlined a number of energy supply and demand scenarios in their publication. Depressingly, even in their “Decarbonisation delivers” scenario where renewables mostly replace fossil fuels, global temperatures would rise by 2.3oC relative to pre-industrial levels by 2100, thus failing to meet the 2015 Paris Agreement target of 2oC. Worse, in their “Hydrocarbon Resilience” scenario, where growing demand for electricity due to AI and wider economic growth is met by burning fossil fuels, global temperatures could be 3.0-4.0oC higher by 2100 [18a].
A world in which global temperatures are 4.0oC higher is likely to see substantial rises in sea levels, lethal heat and humidity, water shortages and food insecurity [20]. This could trigger famine, mass migration and war, and could lead to the end of human civilization as we know it [21].
Conclusion
I am not convinced about the hype about super-intelligent AI wiping out the human race. If it really is so intelligent, it will realise that humans can do things it can’t, and that it needs human beings for its existence. Assuming it has a bias towards self-preservation, it will seek to preserve the human race.
I am more concerned about not-so-intelligent AI doing daft things, or enabling humans with evil intent to hack critical infrastructure or develop new plagues. The threat from AI-enabled bio-terrorism – or accidental release of AI-developed pathogens – is truly existential.
However, we should not lose sight of other existential threats. The threat of nuclear Armageddon was present long before the development of AI, and is growing as more countries like North Korea develop nuclear weapons. Hallucinations by AI embedded in nuclear command and controls systems could increase the risk of mis-calculations with catastrophic consequences.
While nuclear conflict is a possibility, climate change is a certainty. We don’t know how far the climate will warm up, and what impact this will have, but AI energy consumption is likely to increase carbon emissions. To my mind, this is the most likely way for AI to undermine human civilization as we know it, and deserves more attention than fevered dreams of the Terminator.
[1] Anthropic researcher believes more than 10% chance AI 'could kill all humans', BBC, 9th September 2026
[1a] AI Safety Leader Stopped Saving for Retirement Over AI Extinction Fears - Business Insider, 27th October 2026
[1b] While I do not doubt AI systems ability to develop and augment themselves, the dire predictions assume that AI could direct robots to maintain physical infrastructure such as grids and data centres to keep these going in the absence of humans. However, while Elon Musk has promised that Tesla would develop factories powered by humanoids producing humanoids, this has yet to materialise (Optimus (robot) - Wikipedia). By the same token, while Chinese developers are making huge advances in robotics such as humanoids, the latter still appear to be mostly for show (China’s humanoids are dazzling the world. Who will buy them?). To my mind, for some time to come, grids will still rely on humans climbing poles after storms to keep electricity flowing.
Even if they didn’t, and an AI decided to dispense with humans, it would be relatively easy for humans (particularly Americans with their access to guns) to attack transformers and other critical grid infrastructure (see [12] below). A super-intelligent AI would probably settle for influencing and controlling humans through social media.
[2] 1983 Soviet nuclear false alarm incident - Wikipedia
[4] WannaCry ransomware attack - Wikipedia
[4a] Global cyberattack strikes dozens of countries, cripples U.K. hospitals - CBS News, May 12th 2017
[5] Colonial Pipeline ransomware attack - Wikipedia
[6] Why the world’s richest country can’t defend vital infrastructure, Economist, 19th August 2026
[7] Anthropic suspends all access to Mythos model after US government bans foreign nationals use | CNN Business, 13th June 2026.
[8] Is the internet still safe?, Economist, 22nd October 2026.
[9] From IBM and Ponemon Institute’s “Cost of a Data Breach Report 2026: The AI tipping point” (available at Cost of a Data Breach Report 2026 | IBM), page 20, “For the fourth year in a row, phishing was the top attack vector into breached organizations”
[10] Chernobyl disaster - Wikipedia
[11] Wuhan Robotaxi Crash: 100+ Baidu Cars Halt Traffic, IEV China, 8th August 2026
[12] The disaster that could follow from a flash in the sky, Economist, 13th July 2017.
[13a] Black Death - Wikipedia
[14] How AI tools could enable bioterrorism, Economist, May 5th 2026
[15] 1977 Russian flu - Wikipedia
[16] 2026 Russian plague scare - Wikipedia
[17] Mirror life is dangerous. Regulators must act before it becomes real, Economist, 1st October 2026
[18] Exhibit 6 from Global Energy Perspective 2026, McKinsey & Company, September 2026
[18a] Exhibit 11 from Global Energy Perspective 2026, McKinsey & Company, September 2026
[19] Electricity generation - Wikipedia with more up-to-date generation figures available from Yearly Electricity Data | Ember
[20] Comparing climate impacts at 1.5°C, 2°C, 3°C and 4°C | UN Climate Summit News - COP29
[21] I would recommend reading the 2005 book “Collapse: How Societies Choose to Fail or Survive” (2005) by Jared Diamond to get an understanding of how climate change and other environmental factors has led past societies to fall apart, and how exposed modern societies are to these risks.
Send to friend
