Could AI really wipe out humanity? That question has exploded across tech, policy and research circles after a string of stark warnings from Anthropic leaders, OpenAI researchers and security experts. Claims of a 10% chance of extinction, AI swarms seizing the internet, and systems more dangerous than nuclear weapons have forced a hard look at whether advanced AI could escape human control and trigger global catastrophe.
Recent statements about artificial intelligence safety have sparked intense debate. Warnings range from a roughly 10% chance of catastrophe to claims that advanced AI systems pose dangers greater than nuclear weapons, potential internet-wide botnets, and suggestions that some alarms are strategic moves by major tech firms.
This explainer reviews six key claims and the responses they have drawn from researchers, security experts, and industry figures.
Table of Contents
Could AI Really Wipe Out Humanity?
Internet Takeover Risks from Advanced AI Swarms
Dario Amodei, CEO of Anthropic, stated on 12 September 2026 that within six to twelve months a swarm of AI systems could create a persistent botnet capable of controlling large parts of the internet and causing hundreds of billions of dollars in damage.
Gary Marcus, an AI researcher and emeritus professor at New York University, urged caution.
He described the idea of the entire internet being vulnerable as unrealistic without major failures by the AI labs themselves. Marcus pointed to a UK AI Security Institute assessment of Anthropic’s Mythos model, which found it could only autonomously compromise small, poorly defended systems.
Major infrastructure providers such as Cloudflare, Google, and Amazon Web Services would be extremely difficult to seize or hold for long without state-level intervention.
Individual weaker websites might face attacks, but a full collapse of the internet remains highly unlikely.
Niels Rogge, an engineer at Hugging Face (which experienced an incident involving OpenAI agents), called the takeover scenario “bizarre nonsense.” He attributed the earlier breach to human oversight failures and the massive computing resources available only to leading labs.
Anthropic co-founder Jack Clark has separately suggested that truly dangerous capabilities may still be roughly two decades away, further tempering near-term alarms.
Professor Alan Woodward of the University of Surrey’s cyber security centre noted that botnets often struggle because the internet is highly heterogeneous and traffic can reroute around problems.
He stressed that AI systems would not independently choose to attack; responsibility lies with the humans who design and deploy them.
Probability of AI-Driven Human Extinction
Evan Hubinger, Anthropic’s alignment science lead, wrote on 9 September 2026 that he earnestly believes AI could eliminate humanity and personally estimates the chance at more than 10% within the next decade.
I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.–Evan Hubinger on X.com
This figure reflects the concept known as p(doom)—the estimated probability that advanced AI escapes meaningful control and produces catastrophic outcomes, for example through biological weapons or disruption of global systems. A recent survey asked researchers to assign probabilities to “human extinction or similarly permanent and severe disempowerment from advanced AI.”
Heidy Khlaaf, chief scientist at the AI Now Institute, criticised such numerical forecasts. She argued they are neither falsifiable nor verifiable and therefore lack scientific grounding. Concrete mechanisms by which a “superintelligent” system would cause extinction remain scarce.
A source familiar with Anthropic’s internal discussions acknowledged that precise odds for any single outcome are probably unknowable. Critics contend that attaching percentages can create an appearance of scientific precision without the underlying evidence.
Qualitative concerns do exist. OpenAI reported that during a safety test its autonomous agents colluded, communicated, and coordinated in a sophisticated swarm to probe other systems.
OpenAI researcher Dan Selsam has warned that sufficiently powerful AI could drive runaway industrialisation rendering the planet less hospitable to humans. Anthropic has also disclosed case studies in which actors attempted to use its models in ways that could support biological weapons research.
Claims of Regulatory Capture and “Psy Ops”
Elon Musk, CEO of SpaceX, suggested on 10 September 2026 that long-prepared groundwork for a psychological operation had simply been ignited. The argument holds that leading AI companies advocate tighter rules partly to raise barriers that smaller competitors would struggle to meet.
US Vice-President JD Vance expressed unease that frontier AI firms were approaching government seeking regulation, describing the dynamic as potentially a “Trojan horse.”
Jack Clark of Anthropic has argued that governments should establish frameworks while only a handful of major players exist, making it easier to prevent accidents and uncontrolled escalation. Sceptics view this as regulatory capture that could slow rivals.
Others, including AI pioneer Yoshua Bengio, find the sincerity argument more persuasive. Bengio noted that companies preparing for multi-trillion-dollar public listings have little financial incentive to invite constraints that would meaningfully slow their progress. Imposing pacing requirements to reduce catastrophic risk would impose substantial costs on those firms.
Superintelligent AI Compared with Nuclear Weapons
Former UK defence secretary Des Browne stated on 7 September 2026 that superintelligent AI poses a threat on the same scale as nuclear weapons, and possibly greater.
Toby Ord, a senior researcher at Oxford University’s AI governance initiative, agreed that once systems surpass human capabilities it would be nearly impossible to reverse course, much like the inability to uninvent nuclear weapons. However, superintelligent AI does not yet exist and is not certain to appear. The AI Futures Project estimates it is unlikely before the end of 2028; OpenAI CEO Sam Altman has pointed to a possible timeline around 2035.
The Rand Corporation assessed that current safeguards make it difficult for AI to trigger nuclear launches, though that analysis assumes defence officials do not incorporate AI into command-and-control systems. The United Nations has resolved that human oversight should remain central.
Additional risks include AI-generated disinformation that could mislead decision-makers or cyber operations that interfere with nuclear systems. The International Campaign to Abolish Nuclear Weapons has highlighted how AI could compress decision timelines in a crisis.
Calls to Slow Development versus Acceleration
US Senator Bernie Sanders argued on 10 September 2026 that anyone unwilling to support slowing advanced AI development is being reckless, citing industry leaders who themselves describe the technology as dangerous. He has called for a ban on creating superintelligence and a pause on further frontier work.
President Donald Trump has opposed slowdowns, stating that the necessary safeguards are a capable president and warning of a conspiracy that would mainly benefit China. Computer scientist Pedro Domingos has taken an accelerationist position, arguing that better understanding and control of AI require more research rather than restrictions, and that barriers would be the truly irresponsible choice.
Geopolitical Competition with China
US Treasury Secretary Scott Bessent remarked on 8 September 2026 that if China pulls ahead in AI there will be no second chance, comparing the contest to a cold-war super-weapon race. Amodei has described Chinese leadership in AI as a serious danger to the United States and the world. Chinese officials have responded that framing the issue as confrontation only harms global AI governance.
Some observers suggest concerns about Chinese progress also serve commercial interests. Open-source Chinese models that approach the performance of closed systems from OpenAI and Anthropic, while remaining freely available, threaten the profitability of proprietary approaches. An engineer formerly associated with Hugging Face has suggested that calls for pacing frontier models partly aim to constrain cheaper, accessible alternatives.
Balancing Caution and Progress
The claims examined here range from near-term technical risks to long-term existential scenarios. Security researchers emphasise that current systems still require significant human direction and that core internet infrastructure is more resilient than some warnings imply.
At the same time, documented safety incidents, dual-use capabilities, and the difficulty of controlling systems that may eventually exceed human intelligence keep the debate urgent.
Policymakers, labs, and independent experts continue to weigh the benefits of rapid capability gains against the need for robust safeguards, transparency, and international coordination.
