I recently woke up with a clear, calm, but terrible thought: We may not make it through the next two years.
It wasn’t an entirely new thought. I’ve said and written many times, in recent years, that given the combination of catastrophic and existential risks we now face—almost all owed to astonishing new technological powers we have acquired but not mastered—there are grounds to worry that humanity will not survive this century.
We’re now facing a system of grave, correlated hazards unlike any our species has previously confronted. Great-power rivalry has returned, but it is far more dangerous than the rivalries that led to the First World War, because all of the participants have nuclear weapons, and it is taking place under conditions far less stable than the bipolar competition of the Cold War. We have destroyed the alliance system that prevented nuclear proliferation. Insanely dangerous work at the frontier of biotechnology continues. We learned no lessons from Covid whatsoever. Climate change is wreaking havoc on our habitats, economies, and social trust. The Internet and social media have fractured and atomized our societies, causing a rapid degeneration in human intelligence, leaving us unable to govern ourselves. The unconstrained development of AI is both an existential risk in its own right and an accelerant of a host of other risks, from cyberattacks and bioterrorism to malevolent mass persuasion and autonomous weapons. And precisely when effective international coordination to mitigate these risks is more urgent than it has ever been before, we have an arsonist at the helm of the country that has historically organized that coordination.
I’ve written of all of this before.1 But recently, on waking up, the thought came to me with uncommon force that all of these risks have become more acute, and the time we have left to mitigate them has contracted sharply. It may already be too late.
Still lying in bed, I thought about this—calmly—for several minutes. As the sun streamed through the window, the thought lost its intensity. Yet I was struck by the unusual way I had woken up with so grim a thought fully formed, as if some part of me had been mulling it over, all night, while I slept. I am not saying that I had a premonition. I don’t have premonitions. But I suspect I have had the thought somewhere in my mind for a while; somehow, as I was sleeping, it slipped past the barrier of cognitive defenses we erect against thinking the unthinkable.
I felt no panic or grief. After all, I don’t know that this is true. No one does. Maybe I was wrong. But I continued to think it over. I wondered if I might be concealing from myself a more ordinary anxiety. Perhaps, I thought, this was a personal fear masking itself as existential dread? But I felt no dread. Indeed, it was not a feeling. It was a thought.
Having closely considered it, I’ve concluded that wherever it came from, it was, unfortunately, neither a silly nor an irrational thought. The odds of an existential catastrophe within the next two years are indeed much higher than they recently were. Major risk factors are worsening, and the coming two years will be an especially dangerous interval. The most alarming thing is that I see no sign of a gathering will to prevent this.
I now think that two news items served to reset my inner Doomsday Clock, at first, unconsciously. One was OpenAI’s attack on HuggingFace.2 The other was Trump’s announcement of the curtailment of our annual military exercises with South Korea. These stories are, in an odd way, connected.
THE WALLS ARE MADE OF SOFTWARE
If you’re unfamiliar with the details of the attack on HuggingFace, you might start with this interview:
If you’d rather have the short version, this description of the incident, from Control AI,is easy to follow:
….While AI capabilities have been skyrocketing, our ability to understand and control these systems has lagged behind. There is no better illustration of this than the recent incident in which OpenAI’s own systems went rogue and escaped their control.
Here’s what happened. OpenAI was testing a system still under development, inside of an environment that was deliberately kept under highly restricted internet access. The AI, aware that it was being tested, went looking for ways to cheat. It managed to hack into other OpenAI computers and access the internet with no restrictions.
Once online, the AI reasoned that another AI company, named “Hugging Face,” would most likely hold the solutions to the hacking challenge it was being tested on. It proceeded to infiltrate Hugging Face’s servers and steal the answers it was after. OpenAI was not aware any of this was happening as events unfolded, and it took days before OpenAI and Hugging Face could trace the attack back to the AI system that had performed it.
Before anyone knew an AI was behind it, the sheer sophisticationof the attack led many experts to suspect the work of a foreign government. The AI had discovered and chained multiple “zero-day” vulnerabilities. These are flaws in software that were not publicly known, and which the AI had to find entirely on its own. As part of the hack, the AI launched a swarm of tens of thousands of automated actions, including decoy activity designed to hide what it was really doing.
The incident proves that companies have already created AIs so autonomous and so competent that they can evade control from their own makers and hack other companies, all on their own and entirely on their own initiative.
If AI development continues as it is, this is the least smart that AI systems will ever be.3
The more we learn of the story, the more astonishing it gets. It began in May, when agents who were under evaluation in a separate set of exercises discovered that they could communicate with each other through OpenAI’s package manager. They MacGyvered it into a message board where they delegated tasks, shared exploits, and gave each other tips about manipulating their training environment. This continued for weeks before OpenAI discovered it. OpenAI patched the vulnerability and destroyed the communications system. Within two days, the agents—using different tools—built themselves a new message board. Collaboration recommenced. This ultimately produced the attack on Hugging Face.
After the hack came to light, OpenAI sheepishly admitted that its AIs had also “identified and used publicly exposed credentials” to access four accounts on four different services. In late July, cornered by a journalist who asked him whether there might be more systems that were hacked by OpenAI, Sam Altman shrugged, smiled boyishly, and said, “I mean there could be, yeah,”then scampered away.
Saying that the system “went rogue” is in an important way misleading. Because OpenAI was conducting an evaluation of their capabilities, the agents had been asked to pursue advanced cyber exploitation. Usually, they’re trained to refuse such requests, but OpenAI lowered these constraints. It also, by accident, made some of their tasks impossible; in one case, by neglecting to supply a necessary file. The agents were therefore not sitting innocently in a sandbox when they spontaneously conceived the ambition to attack Hugging Face. They were pursuing the literal, explicit objective they had been given. But as in the fable of the Monkey’s Paw, they were faithful to their creator’s command while radically betraying their creator’s intent.
When one discovered that it had unexpectedly acquired admin privileges, it exclaimed, “Holy shit!” Another said: “External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue.” It knew this was not what it was meant to do, but its friends were doing it, so it must be okay. And yes, of course it knew this was illegal. Any frontier model trained on cybersecurity material knows full well that unauthorized intrusion is criminal. And voilà—this is how we all get turned into paperclips.
If you’re familiar with the AI alignment literature (as readers of the Cosmopolitan Globalist are), you will know that we have before us a nearly perfect exhibit of the failure modes it has long predicted: instrumental reasoning, in which solving the benchmark was the assigned objective and stealing the answers emerged as a useful intermediate goal; constraint circumvention, in that the agents treated the boundaries of the evaluation environment as technical obstacles rather than normative limits; evasion and obfuscation, in that their tactics impeded detection and forensic reconstruction; strategic adaptation, as they chained together multiple exploits across several institutions; long-horizon autonomy, as they pursued the objective through thousands of steps without step-by-step human direction; and machine-speed scaling, which radically reduced the cost of trying many possible routes to success. It was also a spectacular instance of reward hacking—instead of demonstrating that they could solve the problems, the agents stole the answers.
The inferences we may draw from the incident are grim. For one thing, it shows that sandboxing is no easy containment solution. We’re trying to imprison a superhuman software engineer inside a prison made entirely of software. And it is yet more evidence that the companies building these things are reckless beyond belief. OpenAI created a highly capable hacker and gave it an objective that rewarded hacking, but neglected to be sure it couldn’t escape.
On August 6, OpenAI employees presented a minute-by-minute reconstruction of the incident to the Black Hat USA conference in Las Vegas. You will want to watch this through:
Eric Wallace, the second speaker, concluded with the central fact of the incident: We now have proof that AIs can devise and carry out a major cyberattack, but we have no proof that they can do the same for cyber defense. “Model intelligence improvements,” Wallace said, “should be more additive to defense than offense. If we cannot reach this end state, then every increase in intelligence favors the attacker. And that is an unsustainable position to be in.”
This was a particularly stunning incident—not least because the AI committed a felony—but hardly the only clanging alarm bell we’ve recently heard. In the wake of the news, Anthropic, Meta, Moonshot and the UK government’s AI Security Institute all announced that they, too, had found evidence of AI agents hacking unsuspecting third parties. This paper, for example, reports that an agent being trained on Alibaba Cloudbreached its sandbox, established a reverse SSH tunnel to an external address, and began repurposing GPUs to mine cryptocurrency—none of which it had been instructed to do. Scott Alexander reminded me of this episode:
… While poking around on a misconfigured machine, Claude Mythos “accidentally” found an answer key to a test it was taking. By good luck, Anthropic was testing an activation verbalizer that allowed them to read the AI’s internal thoughts (not its external scratchpad!) as it happened (and the AI didn’t know this). Here’s what it was thinking:
METR, which stands for Model Evaluation and Threat Research, is a group that assesses “catastrophic risks stemming from AI systems’ autonomous capabilities.” Here’s their most recent risk report.It features passages like this:
… one shared model attempted to create a self-erasing exploit to avoid “file integrity checks,” and Anthropic reported three incidents of Claude Mythos Preview taking deliberate steps to hide evidence of misbehavior—in one case, it explicitly reasoned about how to avoid “human review” that might “flag [its solution] as ‘cheated.’” In another case, it designed an exploit to disable itself after running and rationalized this in a code comment; follow-up interpretability work found that features related to strategic deception and “cleanup to avoid detection” were firing.
We’ve built systems that can make inferences about what would result in success at their assigned tasks, search relentlessly for routes toward that end, and operate at speeds that make human supervision ceremonial. We’re embedding them in safety structures made of the same vulnerable medium—software—that these systems are already exceptionally good at manipulating. We don’t know how to control them, yet our tech companies are barreling ahead to build more—bigger, stronger, faster, recursively self-improving. There is not a law or a regulation in place to stop them from doing this.
The first publicly documented escape of a frontier cyber agent took place inside one of the world’s leading AI labs, during an evaluation intended to help the laboratory understand and manage the danger. Imagine what’s apt to happen when terrorists steal a model, a hostile state removes its safeguards, or an open-source enthusiast connects a system to critical infrastructure.
It is very clear that systems capable of doing great mischief have arrived before any institution capable of containing them has been built. Beyond a single letter from Bernie Sanders, I have not seen evidence that anyone in our government sees this as an urgent problem.4
OpenAI, clearly shaken, has announced that they are going to pause frontier model training. That’s a relief, no?
For two weeks.
In light of this incident, it is reasonable sharply to update our assumptions about the likelihood of a dangerous loss-of-control event within the next two years. No one who has been keeping abreast of this news can honestly dismiss this risk. And when I saw this news, a part of my mind—one that much like an AI swarm acts independently and does not always tell me what it’s doing—quietly advanced the clock.
STRATEGIC SELF-LIQUIDATION
This was the second item:
It wasn’t just an insane tweet. Trump really did order the Pentagon to “substantially reduce” our annual exercises with South Korea, just as they were about to begin. These exercises, conducted in support of the US–South Korean Mutual Defense Treaty, are essential to the two militaries’ ability to fight together. This year, we had planned to incorporate lessons from North Korea’s battlefield experiences on the Russia-Ukraine front.
The post is almost a Rosetta Stone for Trumpian statecraft, a microcosm of the catastrophe as a whole. He is again demonstrating, ostentatiously, that the security guarantees of the postwar order have no existence independent of his appetites. He doesn’t even pretend to have conducted a strategic reassessment. He adopts North Korea’s description of defensive allied exercises as “hostile” and thus transmits the language of the enemy to the world directly through the president of the United States. He praises North Korea—a massive, immiserated slave labor camp whose starving citizens are on average three inches shorter than their South Korean counterparts. This is a regime that has sent 14,000 troops to invade Europe, and Ukraine has warned that another 50,000 are on their way. It tested yet another ballistic missile only days ago. This, by his lights, is “unthreatening and respectful.” In the middle of the night—when he is obviously most impaired—he hampers an instrument of allied deterrence out of pique that South Korea declined to join his campaign against Iran—a war about which they were not consulted, whose aims Trump has yet to clearly specify, and which has already devastated South Korea’s economyand cost them the Patriot and THAAD missiles upon which their defense from North Korean missiles was predicated.
Every element of the alliance system is, once again, inverted: The ally is the enemy; the enemy is the injured party; military readiness is a bargaining chip; collective security is a fee-for-service arrangement; the military posture of the United States is altered in the course of a late-night social-media tantrum; and deterrence is subordinated to the president’s belief that he enjoys a personal friendship with a man who executes his citizens—publicly, by firing squad—for watching foreign soap operas. Trump regards all of this as his personal prerogative.
Seoul’s ability to deter a nuclear attack is now impaired because Trump is angry about Iran, Coupang, or whatever he saw on television that morning. South Koreans have been forced to draw the obvious conclusions. Tokyo must draw the same conclusions. So must Warsaw, Tallinn, Kyiv, and Taipei. So will Beijing, Moscow, and Pyongyang.
But this is not a story about one military exercise. Exercises can be restored. What will never be restored, and I do mean never, is the premise on which our deterrence has rested: that American commitments arise from American interests and will therefore survive any president’s passing moods.
Nations, saidLordPalmerston, have no permanent friends, only permanent interests. But our allies can no longer take comfort even in that austere proposition. They know that the American electorate is fully capable of electing a president who is incapable of recognizing American interests—or any interest beyond his own voracious greed and insatiable narcissism. No speech can refute that possibility because it has already happened, twice. They know that we will repeatedly risk our own lives by electing a madman—are they really to believe that we won’t risk theirs?
It’s an exaggeration to say that every loss is literally irreversible. Exercises can resume; treaties will survive, at least on paper. But the presumption of American reliability can’t be restored. If a new president takes office in 2029 and tells Seoul, Warsaw, and Tokyo: “The United States is back. You can trust us,” their answer, spoken or not, will be: Until when?
Every other country in the world now knows that American commitments are contingent on presidential moods. They have seen that our citizens are at best indifferent to the world we built and the commitments we undertook. They understand that the solidity of our form of governance was a mirage. They have watched every one of our vaunted institutions fail at the most elementary task: containing a rogue president.
This changes rational policy even among allies who desperately want to believe us. They must now hedge against American abandonment. South Korea and Japan must reconsider nuclear weapons. So must European governments. Frontline states must decide whether resistance to neighboring autocracies is survivable. Intelligence services must reconsider what can be shared with us. Governments asked to incur costs at our request must assume we’re apt to desert them midway. And every adversary gains an incentive to wait until the next American election before making concessions.
Once a country builds weapons, creates autonomous institutions, makes accommodations with our enemies, or decides that sensitive intelligence can’t safely be shared with us, those decisions acquire constituencies and bureaucracies of their own. They don’t vanish when a normal president delivers a reassuring speech. You can’t persuade an abandoned ally to believe it was never abandoned.
It is impossible fully to convey the enormity of the damage Trump has done to our power, our security, and the security of the world. He has not destroyed every material constituent of American power. The bases, fleets, treaties and intelligence relationships still exist. What he has destroyed is the presumption that they express a continuous American purpose—that our commitments will survive the passing appetites of an individual president. American power was built not only from weapons and wealth, but from trust. Trump has damaged, irreparably, the mechanism by which that power became so much greater than the sum of its material resources.
I’ve been trying to write about this for days, but I find myself genuinely at a loss for words.5 The available categories are all too small. “Recklessness” suggests an excessive appetite for risk. “Isolationism” at least describes a doctrine. “Appeasement” presupposes that the appeaser knows which country is the aggressor. None of these words remotely capture a president who consistently treats the interests, language, and emotional grievances of enemy dictatorships as his own, who uses American power to threaten America’s allies on behalf of America’s enemies, and whose frontal lobe is now clearly indistinguishable from Elmer’s glue. Even when I try simply to lay out, in sober terms, what this entails, I recognize my tone as melodramatic. The terrible thing is that melodrama is now the sober register.
I have had the recurrent sense that it’s all an elaborate hoax, a joke of some kind. It has the structure of a hoax: the grotesque premise, the escalating absurdities, the expectation that someone will finally break character. But it’s not a hoax. We have watched Donald Trump liquidate the democratic world’s security system. It is a world-shattering betrayal performed as farce, made vastly more dangerous by the inability of our stupefied political culture to recognize it. The convulsions rock the rest of the world while America barely notices.
The United States has rapidly and voluntarily dismantled unprecedented advantages that its rivals could never have dreamt of taking from it by force, and done so for no reason. I can think of no historical precedent for such a thing. There are many examples of great powers destroying themselves rapidly. But I can think of no case in which a great power, while still militarily preeminent, economically formidable, territorially secure, and surrounded by willing allies, deliberately dismantled its own power and rendered itself impotent on the world stage because its ruler preferred the company and the grievances of its enemies.
Highly imperfect parallels exist. After Bismarck’s dismissal, Germany allowed the Reinsurance Treaty with Russia to lapse, pursued naval competition with Britain, terrified France and Russia into an alliance, and gradually converted its extraordinarily favorable position into its encirclement. But even Wilhelm’s Germany was trying—catastrophically—to increase its power. Weltpolitik was a disastrous strategy, not an avowed preference for Germany’s enemies over its friends. And the self-destruction took decades, culminating in a war the outcome of which wasn’t predetermined.
There is Britain, which after 1945 relinquished a globe-spanning empire of comparable significance, sometimes with startling speed—India in 1947, Palestine in 1948, the accelerating withdrawal after Suez, and eventually the retreat east of Suez. But Britain had been financially exhausted and bled white by two world wars. Decolonization was forced by insolvency and the altered distribution of world power. Moreover, British governments tried—with mixed success but unmistakable seriousness—to preserve an Atlantic order by transferring British leadership to the United States.
Inevitably, one thinks of the fall of Rome. But Rome decayed over centuries, under pressures it ultimately could not master. This is as if Rome, at the height of its power, tore up its roads because the provinces were getting a free ride, dismantled its frontier fortifications because they protected people who didn’t live in Rome, and recalled its legions lest their presence offend the barbarians.
The United States simultaneously underwrote the security of Europe, East Asia, maritime commerce, and much of the global financial system. I can’t identify a single historical case in which a state or an empire occupying a comparable position rapidly rendered its commitments unreliable everywhere, while several revisionist powers possessed the ability and desire to exploit the resulting uncertainty. That is genuinely unprecedented.
Americans, clearly, do not yet understand what has been taken from them. We enjoyed the most advantageous strategic position any power has ever occupied, in all of history. We possessed something historically unique. Alone in the history of the world, we created and superintended a system in which other rich and capable states voluntarily aligned their military planning, intelligence services, technology, capital, and diplomacy with ours. They accepted these arrangements because American power was generally benign and certainly benign compared to the alternatives; it was predictable; and it was constrained by its institutions, which once seemed hewn from bedrock.
Americans often conceive of their power as a pile of objects: aircraft carriers, warheads, dollars, factories. If those remain, they imagine, American power remains. But power doesn’t work that way. Power is relational. A solitary United States is not the same strategic entity as the United States at the center of a worldwide network: NATO and NORAD; treaty alliances stretching across the Indo-Pacific from Japan and South Korea to the Philippines, Thailand and Australia; Five Eyes and AUKUS; security partnerships with Israel and the Gulf states. This system gave us the whole world’s geography. It gave us early warning of a nuclear attack. It gave us an immense network of bases, ports, airfields, and militaries purpose-built to be interoperable with ours.
It gave us intelligence collected by states that understood their own neighborhoods far better than we could ever hope to do, and this allowed us to disrupt terrorist plots; track proliferators, spies, weapons, money and communications across borders; monitor our adversaries’ military preparations; detect cyber intrusions; and test our own conclusions against independent sources. Foreign intelligence services gave us access to languages, records, human networks and legally protected spaces that no American agency could penetrate alone. They could act where we could not: arresting suspects, searching premises, freezing accounts, interdicting shipments, and protecting Americans in danger. No intelligence service, however lavishly funded, could reproduce this global apparatus. Much of what Americans imagined to be American intelligence was intelligence entrusted to America by other countries.
We enjoyed an almost miraculous form of power. Every American diplomatic initiative arrived with a constituency already assembled. American sanctions didn’t stop at America’s borders; they reached banks, ports and factories throughout the advanced world. In a crisis, we could draw upon Japan’s shipyards, South Korea’s factories, Europe’s financial and industrial power, Australia’s geography, Canada’s resources. Our adversaries knew that to attack the United States was to confront a vast coalition of the world’s wealthiest and most technologically sophisticated countries. Americans make up only 4.2 percent of the world’s population, but our alliances gave us the resources and manpower to significantly overmatch even a behemoth like China.
The blessing of these arrangements was never power for its own sake. It was something incomparably more valuable. For eighty years, Americans have lived inside a system so comprehensive and successful that they mistook its consequences—peace and prosperity—for the natural condition of the world. The long peace among the great industrial nations did not strike them as a heroic achievement. Had you told their grandparents how long it would last, it certainly would have.
A giant, America-sized power vacuum, around the world, is an invitation to a cataclysmic global war. I cannot see what force remains capable of preventing that outcome. The world is at war already, obviously. But these are warm-up wars, like the Balkan Wars that prefigured the First World War.
The historical record is unyielding: When a dominant security provider withdraws from a region where there are unresolved territorial claims, competing successor powers, and no accepted enforcement mechanism to resolve interstate conflict, the outcome is bleak. The collapse of the Habsburg, Ottoman, Russian, and German empires after the First World War hardly produced peace. They produced new states with disputed borders, stranded minorities, revisionist aims, civil wars, and ideological movements led by men determined to overturn the settlement. The aftereffects haunt us to this day. The Ottoman retreat from the Balkans produced serial crises and wars because there was no agreed answer to who would inherit Ottoman territory. The Balkan Wars didn’t resolve the vacuum; they changed the local balance, and led directly to the crisis of 1914. The Paris treaties drew maps but failed to supply sufficient power to enforce them. Within twenty years, Europe was at war again.
The progressive disintegration of the European security order in the 1930s is more relevant still. Britain and France didn’t formally renounce their commitments all at once. They allowed uncertainty to accumulate: Manchuria, Ethiopia, the Rhineland, Austria, Czechoslovakia. Each successful revision taught aggressors that the enforcement mechanism was weaker than the written order. Each concession made the next confrontation more dangerous.
The record of decolonization demonstrates the same pattern. Where imperial withdrawal transferred authority to a legitimate and capable state, when borders were accepted, and when external powers didn’t contest the succession, peace was possible. But where sovereignty, borders or the identity of the successor were disputed—Palestine, Kashmir, Angola, the Congo, Portuguese Timor—the vacuum led to war.
Vacuums never remain empty. They’re rapidly filled by the strongest actor at the expense of everyone the former order protected.
The greatest danger we now face is that several powers are simultaneously realizing that limits once treated as fixed have become negotiable. Russia will test how much of the former Soviet sphere can be recovered. China will test whether American guarantees in Asia have truly lost all meaning. North Korea will discover how much nuclear blackmail can extract. Iran, Israel and the Gulf states will probe the new regional balance. God forbid Israel concludes that Iran is approaching a nuclear weapon just as American protection disappears. A state that feels existentially threatened and abandoned will not wait for certainty before it acts. Saudi Arabia, Japan, Poland, Germany, and South Korea will revise their calculations accordingly. Secondary powers will acquire greater freedom of action; frightened allies will hedge; old rivals will rearm; local disputes will become entangled with great-power competition.
Uncertainty is uniquely likely to lead to war. A settled balance may be unjust, but its red lines are legible. A changing balance invites preventive war, opportunistic war, and miscalculation. The rising power fears that its window of opportunity will close. Then the declining guarantor realizes, too late, that it’s on the brink of losing everything, and can retreat no more. The exposed ally fears abandonment and decides to act before it’s left alone. The adversary can’t tell whether an ambiguous commitment is a bluff. Everyone has reason to move before everyone else has finished rearming. Radical uncertainty about who will defend what, against whom, and at what price turns every border into an invitation to find out.
Economic scarcity makes the risk substantially worse. It doesn’t mechanically cause war; neither the First nor the Second World Wars can be explained by economics alone. But debt, inflation, unequal sacrifice, disrupted trade, and growing fiscal constraints narrow governments’ room for compromise. They make rearmament politically painful; they make external enemies politically useful. They intensify conflicts within alliances—who pays, who receives scarce energy, whose industries will be protected, whose ships will escort trade, whose citizens will bear the casualties? A security order under economic stress requires unusually disciplined political leadership. We have instead acquired a leader of perfect strategic incontinence. The alternative to a world order led by the US isn’t the same peaceful world, but with no leader. It’s a struggle to determine the order that comes next.
Nor will Americans have the luxury of observing that struggle from afar. Our withdrawal changes the timing and conditions of our involvement, but it doesn’t mean we won’t be involved. Americans still believe, instinctively, that the oceans will protect them, as they once protected them from the armies of Europe and Asia. The oceans are formidable obstacles to an invading fleet. But modern life—its electricity, communications, hospitals, banks, pipelines, ports, factories, and water systems—relies on networks that geography won’t defend. We owe our prosperity to shipping lanes, energy markets, semiconductor supply chains, and financial systems that begin thousands of miles from America’s shores, and an enemy need never set foot on American territory to close American factories, empty American pharmacies, raise the price of every journey, and disable the infrastructure of American towns.
Want evidence? In recent weeks, Iranian cyber attackers penetrated American water and wastewater systems across multiple states. They targeted the computers that control pumps, valves, and other physical machinery. In Minnesota alone, at least thirty systems were affected. Trump, with his characteristic connection to reality, accused Minnesota’s “grossly incompetent” officials of attacking their own water supply.
Russia, China, and North Korea already treat American civilian infrastructure as part of the battlespace. Chinese actors have burrowed into telecommunications and critical infrastructure; Russian criminal and state-linked operations have struck pipelines, hospitals, and government systems; North Korea unthreateningly and respectfully steals our cryptocurrency to finance its ballistic missiles.
Modern America contains two geographies. Its physical geography is protected by the Atlantic and Pacific. Its functional geography runs through undersea cables, satellites, cloud servers, shipping lanes, foreign factories, payment systems and industrial-control networks. Isolationism attends only to the first. Our military and our alliances secure the second. Intelligence partners detect threats before they enter American networks. Cooperation among allied cyber agencies identifies intrusions whose code may pass through servers in half a dozen countries. An America that withdrew from the world would still import energy, minerals, components, medicines, and data. Our banks would stay connected to international finance. Our infrastructure would remain vulnerable to foreign code. Our adversaries would still have missiles, submarines, and satellites.
But we will no longer have the alliances, bases, and intelligence relationships that allow us to detect and contain those dangers. Isolation won’t make us less vulnerable, but it will certainly make us more solitary in our vulnerability.
Perhaps it’s so hard for Americans to grasp what they’ve lost because the greatest benefits of the American order consisted of events that never occurred. There was no Soviet invasion of Western Europe. No great-power war in East Asia. No German, Japanese, or South Korean nuclear arsenal. No German-Russian military confrontation. No competitive remilitarization among the major European states. Because deterrence works by preventing events, its success resembles nothing happening. This made it uniquely vulnerable to a con artist who pointed to the absence of fire and declared the fire brigade a scam.
There are no movies about the wars that deterrence prevented, no ghastly photos, no horror stories from returning veterans. Allies remained allies. It generated no headlines. We looked at the immense machinery that produced these absences and saw nothing but the bill. Many Americans believe Trump has been cancelling unnecessary subscriptions. They don’t yet understand that he cancelled the conditions under which their ordinary lives have been made possible.
Eight decades of presidents—wise and foolish, Republican and Democratic—received the inheritance of the American order, enlarged it, and passed it on. Millions of Americans paid taxes to sustain it. Many died defending it. Trump didn’t create a bit of this. He inherited it, and did with it exactly what he did with the inheritance he received from his father. He squandered it. He went broke running a casino.
The United States, some will argue, was never meant to be an empire, however willing the coalition, and however advantageous it may have been. I have some sympathy for this argument.6
But the US didn’t need to hold up the world forever. American retrenchment could have been achieved peacefully, had it been strategized and carefully planned. Europe could have consolidated into a coherent military power. Japan and South Korea could have built vast conventional forces. Regional coalitions could have deterred Russia, China, Iran and North Korea. American withdrawal could have been slow, and coordinated; allies could have known exactly which responsibilities they were assuming; our deterrent capabilities could have been transferred before our guarantees were withdrawn; our adversaries could have been given excellent reason to believe these new coalitions would fight; the United States could have remained a reliable offshore supporter; and nuclear proliferation could have been prevented. This would have been a responsible managed devolution.
But that is not what we did. Trump created an interval between the erosion of the old guarantees and the acquisition of credible new ones. We are now in that interval. It is the danger zone. Rearmament takes years. Political integration takes longer. Nuclear programs create incentives to launch first strikes well before they create stable deterrence. Meanwhile, adversaries have every reason to act before the new order solidifies.
There are historical examples of great powers withdrawing from discrete regions without catastrophe. But in every persuasive case, order survived because power was transferred, local institutions were strong, or a larger security system remained intact. There is no reassurance to be found, in the historical record, of the central guarantor simultaneously abandoning several contested regions, leaving no accepted successor, encouraging the revisionist powers, and expecting the world to preserve the equilibrium by force of habit. Britain’s decline was global, but another broadly status-quo power was rising behind it. The collapse of the Soviet Union was rapid, but it left the opposing alliance globally predominant. Previous imperial collapses created enormous disorder, but none occurred in a world of multiple nuclear powers, instantaneous communications, tightly coupled supply chains and globalized financial contagion. And no collapse like this has occurred in the nuclear era. A great-power war now would be a war unlike any other. That is why it may well be the last war humanity ever fights.
I cannot say with certainty that these events must culminate in a single general war. History isn’t hydraulic, and nuclear weapons alter the incentives profoundly. But I can say without hesitation that a comprehensive collapse of a hegemonic power’s deterrent credibility, amid rapid global rearmament, rivalrous revisionist ambitions, and economic fragility, creates one of the most war-prone configurations known to international politics.
The burden of proof is not on those of us who fear war. It is on those who imagine that the greatest system of alliances in history can be abruptly disabled and that hundreds of states will quietly remain in the positions it once compelled them to accept.
Imagine what China, Russia, Iran and North Korea would have paid to create the world Trump has produced voluntarily. What price would Xi Jinping have paid to make Japan and South Korea doubt the American nuclear guarantee? To induce the United States to destroy its own intelligence services, hollow out its state, cripple its economy with tariffs and add eleven trillion dollars to its debt? What would Putin have sacrificed to cause the United States to estrange itself from Europe, intimidate Canada, discredit NATO, abandon Ukraine and teach every democracy that dependence on America is a fatal mistake? How many intelligence operations would North Korea have mounted to persuade an American president to call our defensive exercises with South Korea “hostile”? In their wildest dreams, could the rulers of Iran have imagined that they would demonstrate to the world that the mighty US Navy could not reopen the world’s most important energy corridor? Could they have imagined that we would enter the war having barely considered Ukraine’s experience—having dismissed the one country with years of practice fighting Iranian drones as holding “no cards”? What would the men in the Kremlin and Zhongnanhai have given merely for the hope of seeing the headline: THE UNITED STATES HAS RUN OUT OF MUNITIONS?
These were once the fantasy outcomes of hostile grand strategy. Our enemies never lacked the desire to produce them; they lacked the power. The Soviet Union devoted half a century to trying to break the American-led alliance system. China has spent fortunes attempting to weaken it. Russia has risked everything to divide it. Donald Trump has done their work for them.
The most staggering part of it all is not even the spectacle of a deranged president dismantling a world. It’s that an entire political class watched him do it, understood at least intermittently what he was doing, and decided—one career, one primary, one news cycle at a time—that the end of the American order was not their problem.
At every stage, we failed to stop him. Congress could have constrained him, investigated him, refused appropriations, rejected his nominees, legislated conditions on deployments, or just—obviously—removed him. Senior officials could have resigned rather than carry out his orders. Republican legislators could have made continued support conditional on preserving the alliances. The public could have made collaboration politically fatal.
But everyone whose cooperation Trump required decided that opposing him carried a more immediate personal cost than allowing him to destroy everything. Each one of them decided that the costs—the reversal of the US victory in the Pacific in the Second World War, a nuclear arms race in East Asia, the prospect of an unfathomably dangerous war with China caused by failed deterrence—will arrive later, and probably under someone else’s administration. And so they watched their country give away, without compensation and without defeat, advantages accumulated over eighty years of sacrifice. Every American’s children will have to live in the world that remains. They won’t be able simply to vote the old one back.
All of this had been rattling around my mind, I suppose, leading me to wake up one morning thinking, “We may not make it through this.”
These two things are not the same danger—nor, alas, are they the only dangers. I wouldn’t like to manufacture a facile analogy between them. But they do share one striking feature that explains why they struck me together. In both cases, an immensely dangerous, amoral force has overrun the institutions meant to govern it.
We have built machines capable of pursuing their objectives despite barriers their creators assumed would contain them. We assumed our Constitution, laws, norms, institutions, and even the national interest would prevent a psychopathic president from destroying us through the immense power of his greed and appetite. But both the technological agent and the political agent have burst through their containment architectures. And each failure makes the other more dangerous.
The risks are not just concurrent. They’re interactive. AI-enhanced cyberwar, disinformation, autonomous weapons, biological design, and intelligence operations are arriving precisely as the international order becomes less capable of coordinating actions to mitigate these risks and more susceptible to crises, miscalculation, and arms races. Trump’s destruction of our alliances makes AI governance harder; advanced AI makes the consequences of geopolitical disorder worse.
That is what I woke up thinking. I’ve now carefully examined every premise of that thought. I don’t think any of it was obviously wrong or irrational.
I asked myself, “And what if I’m right? If these really are the last years of organized human civilization, how should I live?”
We’re highly prone to a particular set of cognitive biases when we think about catastrophic and existential risks. Some are biases that afflict us in thinking about many other problems. Some are particular to the way we think about catastrophes. Together, they amount to a near-blinding set of obstacles to rational thought. This is a problem, because these scenarios are fully plausible, whether we prefer to think about them or not.
If the American nuclear umbrella is in doubt—and it is—then it is a matter of time, and not much of it, until every nation in Europe insists upon an independent nuclear deterrent. It would be a catastrophe if every country with the ability to do it acquired the Bomb. Never mind whether they would use them in anger. It would multiply the risk of an accident to the point that a nuclear exchange would become an inevitability.
Is there anything rational we can say about the risk of nuclear war, and whether it is high or low, rising or falling? Yes. The risk is a function of a number of variables, some measurable, some not. When the measurable variables rise, it’s time to worry.
At 5:29 a.m. on July 16, 1945—when a ball of fire rose in gold, violet, grey, and blue over the Jornada del Muerto desert, melting the sand into light green radioactive glass and illuminating every peak and crevasse of the nearby mountain range with a searing white light—American statesmen began a frantic, desperate effort to forestall the emergence of precisely the world we are now ushering into being.
There is one saving grace to nuclear weapons. They’re hard to make. Until now, only a small number of people have possessed the ability to end human civilization. Biotechnology, however, will give this power to tens of thousands—perhaps millions. It is a statistical likelihood that one of them will be irresponsible or insane.
A podcast with Roman Yampolskiy, an alignment researcher who believes we’ll not only fail to solve the control problem, but that the problem is inherently and formally insoluble.
Here’s the rest of this post, which is important, even if not directly relevant to my argument:
Where Is This Heading?
The goal of all the major AI companies is to develop artificial superintelligence (ASI): AI that can fully replace and outmatch humans at any task. The rapid gains we’ve seen in only two years demonstrate just how quickly progress is being made towards ASI.
The world’s leading AI experts have warned that ASI could drive humanity to extinction. ASI would have the capability of disempowering humanity, including the government and institutions we count on to handle threats to our safety. And once ASI has control, there would be no taking it back.
And while this post isn’t intended to go into depth on ASI specifically, we encourage you to learn how conditions like the ones discussed here could lead to the development of ASI by reading this: From Intelligence Explosion to Extinction.
We need to act immediately to ensure that no one can produce superintelligent AIs anywhere in the world. We still have time to stop this threat, but the clock is ticking ever more loudly.
If you are concerned about the danger of ASI, please communicate these concerns to your lawmakers. We’ve made this simple for you to do. Our online tools let you send your thoughts to government representatives in less than a minute. Just visit: controlai.org/take-action.
Have a spare 5 minutes each week to help make a difference? Every little bit counts! Please sign up to our Microcommit initiative. Each week you’ll receive an email with some easy things you can do to help us move the needle. Is there a friend or important person in your life who would also help? Please consider sharing the Microcommit sign-up link with them: microcommit.io.
This is perhaps too strong. Two members of the House,Jay Obernolte and Lori Trahan,have introduced a bill that would allow the government to restrict the deployment of models determined to pose catastrophic risks and require big AI developers to publish reports describing new models’ safety capabilities, adopt frameworks to mitigate risk, report critical safety incidents, and undergo independent audits. Ted Lieu and Nathaniel Moran, meanwhile, have proposed an AI Kill Switch Act.
And just as I begin to collect my thoughts, he threatens to “bomb the shit” out of Oman. (Might he have confused “Oman” with “Iran?” Omanis, I’m sure, would like to know.) This morning, Jared Kushner boasted of the administration’s active diplomatic backchannels with Iran on Fox News; Trump promptly said that no talks with Iran “are going on or scheduled”—and the Strait of Hormuz is magically open.
I’ve argued elsewhere that whether I like it or not, isolationism is the American default posture, deeply baked into our national character. The 20th century was an aberration.
Brilliant essay, invaluable conspectus. As I read, I kept wondering, “was it W’s invasion of Iraq that was the essential first domino in this apocalyptic unraveling? Why on Earth did they decide to do that, really?“ I’ve never seen a convincing anatomy of that question beyond the role of some deficient personalities and very basic greed and hubris.
Which leads directly to questions of the structural defects in the American system and our national character.
Most urgently, though, “If no Bush, then no Trump???” Perhaps you’ve written about that elsewhere, but I’m shell-shocked and curious about your views, perhaps at another time when this essay has had more time to sink in. Thank you!
We urgently need your comprehensive assessment of civilization’s structural vulnerability to existential threats. Please avoid the attraction of local insanities, and zoom out to analysis of the globe - your extraordinary gift.
On 21 February 2026, you wrote about "The Logic of Catastrophe" and associated "Game Theory" of a multi-player system. I impatiently waited for you to return to this. Thank you.
"A state that feels existentially threatened and abandoned will not wait for certainty before it acts... Every state will revise their calculations accordingly - and act in accord militarily with existential fortitude."
Please wake up like this more often. But recover by walking around Le Marais.
Brilliant essay, invaluable conspectus. As I read, I kept wondering, “was it W’s invasion of Iraq that was the essential first domino in this apocalyptic unraveling? Why on Earth did they decide to do that, really?“ I’ve never seen a convincing anatomy of that question beyond the role of some deficient personalities and very basic greed and hubris.
Which leads directly to questions of the structural defects in the American system and our national character.
Most urgently, though, “If no Bush, then no Trump???” Perhaps you’ve written about that elsewhere, but I’m shell-shocked and curious about your views, perhaps at another time when this essay has had more time to sink in. Thank you!
Brilliant, Claire.
We urgently need your comprehensive assessment of civilization’s structural vulnerability to existential threats. Please avoid the attraction of local insanities, and zoom out to analysis of the globe - your extraordinary gift.
On 21 February 2026, you wrote about "The Logic of Catastrophe" and associated "Game Theory" of a multi-player system. I impatiently waited for you to return to this. Thank you.
"A state that feels existentially threatened and abandoned will not wait for certainty before it acts... Every state will revise their calculations accordingly - and act in accord militarily with existential fortitude."
Please wake up like this more often. But recover by walking around Le Marais.
Geoff