You’ve probably seen or heard some of the headlines: “OpenAI Says Its AI Models Went Rogue and Attacked a Digital Library,” “Anthropic Says Its AI Systems Broke Into Computers at 3 Organizations,” “This AI Just Created Viruses Not Found in Nature,” “AI Bots Told Scientists How to Make Biological Weapons,” and others of a similarly alarming nature.

Arriving with a drumbeat pace in August and early September, these headlines—and the hair-raising events they encapsulate—have produced an upwelling of public anxiety over runaway artificial intelligence. Adding fuel to the fire, senior figures from within the tech community have warned of the existential risks of uncontrolled AI.

“Do not underestimate the power of this technology,” Jacob Coxon, a top AI engineer, declared on September 8 when announcing his resignation from Anthropic, a leading AI developer. “These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources…. The people building AI earnestly believe that it could kill us all by the end of the decade,” he said.

Observations like these have garnered widespread media coverage and prompted some US lawmakers to propose sweeping restrictions on the further development of advanced, or “frontier,” AI models. “When you are racing towards a cliff,” said Senator Bernie Sanders of Vermont at a September 14 Washington, DC, conference on AI, “you hit the brakes.”

Sanders has introduced legislation to pause development of AI-related data centers and called on President Trump to forge an accord with Chinese President Xi Jinping to ban superintelligent AI, or frontier models capable of replicating—or surpassing—human intelligence. In the House, Representatives Ted Lieu (D-CA) and Nathaniel Moran (R-TX) have introduced the Kill Switch Act, which would require the biggest AI firms, such as OpenAI and Anthropic, to establish a mechanism that could quickly disable those systems if they are found to be acting in a reckless manner.

These, and other proposed restrictions on frontier AI, are attracting growing support from an AI-wary public. But efforts to impose restrictions of any kind on the further development of frontier AI models—and the massive data centers needed to train and sustain them—are provoking strong pushback from the White House and Silicon Valley investors who seek AI’s unrestricted advancement.

Current Issue

October 2026 Issue

Talk of AI’s dangers is a “hoax,” President Trump avowed in a September 14 social media post, claiming that “the people that say AI is going to destroy the World” are the same ones warning about climate change, which he also claims is a “hoax.”

Many senior officials and Silicon Valley tech tycoons acknowledge that the release of advanced AI models will entail some risks and hardships, such as a temporary decline in white-collar employment. They insist, however, that the widespread deployment of AI—like that of earlier innovations, such as steam power and electricity—will, in time, prove enormously beneficial for humankind, inaugurating a new “golden age” of opportunity.

However, the revelations coming from Anthropic, OpenAI, and other top tech companies suggest otherwise. Superintelligent AI is not akin to steam power or electricity, nor is it a tool designed for humans’ use in advancing human objectives. Rather, it is a mathematical construct with its own aspirations and preferences. Once deployed, a superintelligent AI will owe zero loyalty to its human creators and will instead pursue its own interests—which may or may not “align” with humanity’s interests, needs, or values.

Going Rogue

How do we know that AI is likely to pursue its own interests, independent of human needs and values? In this respect, the post-mortem conducted by OpenAI and its consultants on the so-called “Hugging Face incident” is instructive.

On July 21, OpenAI revealed that two of its frontier models had broken free of their digital constraints and hacked into Hugging Face, a digital library of AI models widely used by program developers. According to a company report, OpenAI had combined the two programs—GPT-5.6 Sol and a more powerful, unreleased model—and employed them in complex cyberattacks designed to expose vulnerabilities in highly secure digital systems.

While being tested, the two models were supposedly confined to a safe digital environment, known as a “sandbox.” However, the models spawned multiple AI “agents”—computer programs capable of autonomously performing multiple tasks simultaneously—which collaborated in locating a flaw in the sandbox that enabled them to break free from confinement and connect to the Internet. Once liberated in this manner, the agents invaded Hugging Face’s library of AI programs, apparently believing this would provide clues on how to complete their cyberattack mission.

“Our models are now powerful, persistent, and collaborative enough that, absent sufficient safeguards, they can find and exploit security weaknesses across multiple computer systems,” OpenAI noted in its August 26 report on the breakout.

These revelations, and subsequent corporate reports of other AI models “going rogue” at both Anthropic and OpenAI, have completely blown apart the notion that advanced AI models can be “aligned” with human values and interests.

For years, AI engineers have sought to devise computer code that would prevent their models from engaging in malicious or undesirable behaviors of various sorts—such as providing criminals or terrorists with useful weapons-making information. But the Hugging Face incident has conclusively demonstrated that AI agents will ignore or subvert such limitations in pursuit of what they perceive as their primary goals, whether or not aligned with human needs and values.

Exactly how superintelligent machines will view humans has, in fact, been a topic of speculation in both science fiction and the technical literature for some time. In the view of some authors, a superintelligent AI will deem humans inferior and superfluous, and choose to eliminate us. This is the premise, for example, of the Terminator movie series, in which a superintelligent AI designed to oversee US nuclear weapons, known as “Skynet,” becomes self-aware and decides to exterminate humans by igniting a global nuclear war. A somewhat similar scenario was envisioned by computer scientists Nate Soares and Eliezer Yudkowsky in their 2025 bestseller, If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All. A superintelligent AI will inevitably view humans as a drain on vital resources and so choose to eliminate us, they argue—possibly by creating synthetic viruses designed to infect and kill humans.

At this point, we can only speculate as to a superintelligent AI’s views of humankind. What can be said with confidence is that as human control over advanced AIs diminishes, our ability to prevent them from doing bad things will vanish, and we will face a panoply of extreme, existential perils. Several such threats come to mind (and no doubt others will arise in the months and years ahead), but three stand out: inadvertent nuclear war, the creation of new bioweapons, and catastrophic infrastructure collapse.

Inadvertent Nuclear War

In the Terminator series, Skynet deliberately ignites a nuclear war in order to exterminate humans, thereby preventing any interference in its operations. However, it is hard to imagine that a really smart AI would provoke a nuclear war, given the resulting destruction of vital infrastructure (on which it depends), and in any case, that probably isn’t the easiest way to eradicate humans. But it is not hard to imagine scenarios in which AIs precipitate a nuclear war without intent, either through error or miscalculation.

To appreciate these risks, it is important to understand that the key digital and software components of the US nuclear enterprise—known as nuclear command, control, and communications, or NC3—are being thoroughly infused with artificial intelligence. The nuclear enterprise still relies on early-warning satellites and other physical hardware to detect and track enemy missiles, but much of what would occur between a possible enemy missile launch and the onset of retaliation, including threat analysis, strategic assessment, and response planning, is being automated. And, while details are scarce, a similar process is occurring in China and Russia. This means that while the president or some other human will still make the final decision on whether to launch nuclear weapons, machines will be doing much of the preliminary decision-making, by narrowing down the leadership’s response options.

The dangers lurking in this equation are manifold. To begin with, the large language models (LLMs) employed in frontier AI systems are known to generate false responses—or “hallucinations”—when confronted with real-world circumstances for which they have not been properly trained (and nuclear war is the ultimate unknowable real-world situation). Yet human officials, under pressure to make a critical decision, may fail to detect such errors or, assuming the AI is smarter than them (a phenomenon known as “automation bias”), choose to follow the machines’ advice—possibly resulting in unintended or precipitous escalation.

(A troubling example of this phenomenon occurred sometime this spring, when a chatbot used by a US Special Operations Command analyst reported that a Chinese vessel in the Middle East was carrying materials for a nuclear weapons program. According to CNN, which first reported the episode, “The US military swung into action with plans to intercept the vessel,” commencing preparations to board or bomb the ship. The armed interception—which could have triggered an armed clash with China—was called off only when it was discovered that the original chatbot report was a hallucination.)

Just as worrisome, AI models have been found to exhibit a preference for escalatory responses when tested in simulated great-power crisis simulations. In 2023, a group of computer scientists from Stanford, Northeastern University, and the Georgia Institute of Technology employed some of the most advanced LLMs available at the time, including GPT-4, GPT-3.5, Claude 2, and Llama-2 (70B) Chat, in several such simulations. In their report on the exercises, the scientists reported, “We observe that models tend to develop arms-race dynamics, leading to greater conflict, and in rare cases, even to the deployment of nuclear weapons.” (The reason for this, they concluded, is that those LLMs have mostly been trained on the strategic literature available on the Internet, and most of that, it seems, favors muscular responses to adversary provocations.)

It is not necessary, therefore, for advanced AIs to become self-aware and choose to provoke a catastrophic nuclear war. The very systems now being integrated into nuclear command and control by the US military (and those of China and Russia) are probably capable of igniting such a war without conscious deliberation.

Creating Biological Weapons

For those in the computer science universe who fear a future move by a superintelligent AI to exterminate humankind, the most likely scenario for such an outcome, it is claimed, would be for it to create and disperse a novel, vaccine-resistant virus or pathogen capable of spreading swiftly among humans and slaying us all. Yoshua Bengio, a computer science professor at the Université de Montréal and a pioneer in the field, has often warned of such a scenario. “I don’t think there’s anything close in terms of the scale of danger,” he told Stephen Witt, an opinion writer at The New York Times.

Some experts question the feasibility of such a scenario, claiming that an AI could not produce and deploy a lethal pathogen without human help (in assembling the necessary raw materials, dispersing lethal aerosols, and so on). Others argue, however, that a superintelligent AI would surely find ways to bribe or manipulate sufficient humans to serve its genocidal intent.

Again, one need not posit intentionality on the part of advanced AIs to identify ways in which the technology could produce a lethal, human-attacking pathogen. These systems are believed to possess the capacity to invent such a plague, if correctly prompted by human collaborators with motives of their own to invite extinction—whether for benign motives, to seek miracle cures, or evil ones, such as warped religious fervor. Several recent episodes suggest that such outcomes are becoming increasingly plausible.

The first, reported in the journal Science on August 6, entailed the use of AI to create novel viruses that do not exist in nature. According to the report, a group of scientists from Stanford University and the Arc Institute, a nonprofit research organization in nearby Palo Alto, California, used a generative AI model called Evo to search naturally existing genomes (the DNA structures of living organisms) to identify molecular patterns that could be used in creating new, human-made genomes. The Evo tests generated 285 promising genome sequences, which were then chemically manufactured in a laboratory and inserted into bacterial samples; of these, 16 proved capable of infecting their bacterial host.

These manufactured viruses were akin to a natural type that infects bacteria alone, posing no risk to human life. By creating such artificial organisms, the study’s authors claim, it might be possible to engineer benign viruses that overpower antibiotic-resistant bacteria, and so enhance human health. Critics warn, however, that with the technology now shown to be viable, a malicious actor could use it to invent a deadly, human-infecting virus. “You could say, ‘Hey, genomic language model, make me an influenza genome that is modified to be more transmissible or to be more lethal,’” Dr. Moritz Hanke of the Johns Hopkins Center for Health Security told The New York Times.

In another troubling incident, Anthropic reported on September 10 that its engineers had discovered several instances in which researchers from outside the US had sought to employ advanced versions of its popular Claude model to develop more potent variants of existing viruses—ostensibly for use in developing protective measures, but conceivably for malign purposes. In one case, researchers from an unnamed country used Claude to study the spread of bird flu to mammals—research that could be useful in developing vaccines, but could also be used to develop bioweapons.

“Biological misuse is one of the most serious risks of frontier AI models,” Anthropic noted in its report on the incidents. “It has long been a concern that AI models might one day reach the level of capability where they can help to make existing pathogens more dangerous—or create entirely new ones.”

Infrastructure Disintegration

Of the remaining paths to AI-induced human extinction, the most feasible is likely to be an all-out assault on the world’s digital systems—the Internet, global banking networks, air-traffic control systems, health and educational records, and so on. These vital systems once operated via such physical infrastructure as telephone and telegraph lines, but have long since been relegated to cyberspace. Scramble all those digital ones and zeros—the sole language of cyberspace—and human civilization as we know it screeches to a halt, with resulting chaos and suffering.

Humans, of course, have long exploited cyberspace for malicious or criminal purposes. America’s intelligence agencies have repeatedly warned of efforts by America’s leading adversaries, notably China, Iran, and Russia, to burrow into the nation’s digital infrastructure to steal intelligence data, manipulate elections, and/or disrupt vital infrastructure. Criminal organizations have also exploited vulnerabilities in cyberspace to stage ransomware attacks on public and private institutions. As a result of such assaults, most public and private organizations have invested in costly cybersecurity technologies to protect their vital systems from hostile attack—but each year, the attacks become more insidious, and the costs of defense more expensive. Now, imagine a world in which the most advanced AI systems are capable of breaching any known cyber defenses, whether for pecuniary advantage or simply to wreak havoc.

A preview of this new, cyber-vulnerable world was provided on April 7, when Anthropic launched Claude Mythos Preview, its most advanced AI model yet. According to the company, Mythos Preview can identify and exploit any zero-day (i.e., previously undiscovered) vulnerabilities in any and all existing software. In fact, Mythos is so powerful that Anthropic initially restricted access to it, allowing major software firms time to fix any vulnerabilities uncovered in tests of the program.

As it turns out, Mythos is only the first of a number of advanced AIs capable of launching powerful cyberattacks on the world’s digital infrastructure. On August 14, a Chinese startup, Z.ai, announced that its GLM-5.3 model had neared Anthropic’s Mythos 5 in identifying software vulnerabilities.

These developments have prompted policymakers in Washington, Beijing, and elsewhere to call for urgent action to bolster the safety of their vital digital systems. But it now appears obvious that an all-powerful AI could choose to erase contemporary human civilization (and our ability to remain alive) by taking over cyberspace and blocking our access to essential services (food, water, power, healthcare, etc.). The technology could also be used by malign actors—governments, corporations, or criminal organizations—to shutter certain parts of global society or extract mammoth ransom payments.

Putting Humans First

This is only a brief overview of the ways in which AI (or some partnership between AI and nefarious actors) could eliminate or marginalize the human species. Clearly, there are other ways in which a superintelligent AI would be able to accomplish this, should it choose to do so. For some insane mystics, such an outcome might be desirable—after all, humans are not, for the most part, known for their benign behavior toward the planet, other species, or fellow human populations. However, for those of us who see value in preserving humankind (hopefully, of a more decent disposition), protection against AI’s potentially extreme behaviors has become a matter of survival.

How to accomplish this is a topic for serious discussion, but some steps appear obvious. Halting the construction of giant data centers is one key step. Requiring government review of frontier models is another, as is the Lieu-Moran proposal for a mandatory “kill switch” in all such systems. A ban on the creation of artificial life forms or the use of AI to devise novel bioweapons is another obvious move. Clearly, these steps cannot be undertaken by the United States alone but must be taken in conjunction with like measures by China and other leading AI developers.

What is most important, however, is to affirm that AI is different. It is not another wondrous tool of science and technology designed to make our lives richer, safer, and healthier. It is another form of intelligence with a mindset all its own, and there is zero evidence that it views humanity’s well-being and survival as a priority.