In the wake of Operation Epic Fury, both supporters and critics of the president have described the joint U.S.- Israeli bombing campaigns as “exquisite.” Iran’s supreme leader, in addition to vast swaths of the security cabinet and IRGC command were wiped out at the same time that Iranian missile production capabilities were severely degraded. The intelligence was, in some senses, so captivating, so advanced, and so hyper targeted, that military experts on both sides of the political spectrum set aside the liabilities of escalating a hot war to celebrate the god-like perfection of its initiation. But if this new intelligence apparatus, powered by AI and irresistible to America’s top commanders is hastily integrated into the nuclear arsenal, the fallout will be anything but “exquisite.”
An examination of government documents, private sector contracting records, and the little noticed statements of military commanders suggests that the same artificial intelligence that allowed frictionless decapitation in Iran is now coming to the nuclear arsenal—with potentially world-altering consequences. While much noise has been made about safeguarding the nuclear command from AI, with constant reassurances of “human-in-the-loop” safeguards, a different escalatory threat has fallen by the wayside: left of launch operations.
With worst-case scenarios of nuclear engagement, most people think of Strangelove-esque military planners fomenting support for a doomsday machine. During the Cold War, near-misses occurred with terrifying frequency, such as the occasion when a flock of geese was mistaken for a soviet nuclear bombing campaign. Most of these almost catastrophic mistakes revolved around mistaking things for missiles that had already been launched.
With the integration of AI into the nuclear command and control infrastructure, escalation may soon begin on the ground, before the launch codes have been entered and the bunkers sealed. This new doctrine is known as “left of launch” and AI is increasingly being integrated into the systems used to predict when a nuclear weapon is being launched, as well as the assets that could be degraded to prevent a first strike.
As we have seen time and time again, the frictionless intelligence that led to a perfect exfiltration in Venezuela, or the targeted killings in Iran, may soon grease the wheels of preemptive strikes on nuclear capabilities, an escalation into untested terrain for both artificial intelligence, and humanity.
What Is “left of launch?” The first public use of the term appears to be a 2014 memo between Army and Navy chiefs discussing the need for new technologies for U.S. missile defense. That memo states that “Now is the opportunity to develop a long-term approach that addresses homeland missile defense and regional missile defense priorities—a holistic approach that is more sustainable and cost effective, incorporating ‘left-of-launch’ and other non-kinetic means of defense. The proposed strategy would serve as the capstone for the Department to balance priorities, inform resourcing decisions, and restore our strategic flexibility.”
This memo marked a strategic shift, putting an increasing focus on the idea of stopping adversaries’ nuclear missiles before they’re launched, a tactic that proved effective during the Obama era for degrading North Korea’s ballistic missile tests. Since the 2014 memo, cyber attacks and sabotage have been added to preemptive air or missile strikes on foreign missile launchers and facilities under the umbrella of “left of launch.”
Five years after the term “left of launch” began cropping up inside the Department of Defense, the Trump administration commissioned and released a review of American “policies, strategies, and capabilities…to counter the expanding missile threats posed by rogue states and revisionist powers.” The report called for an escalatory build-out of new capabilities. Among those was a new framework for advancing the notion that preemptive action is actually just another form of defense.
As Laura Grego of the Union of Concerned Scientists wrote at the time, the report increased the escalation cycle with America’s enemies by “bringing attack operations into the overall missile defense posture as a triad along with active defenses, such as interceptors, and passive defenses, such as hardening and dispersal of potential missile targets.” This, according to Grego, created what amounts to “a kinetic version of left of launch” that no longer reflects traditional deterrence but instead is a first-strike capability.
Grego also pointed out that the report’s interpretation by other nuclear powers, namely Russia and China, will not abide by America’s framing that new satellites and hypersonic ballistic missiles are for defensive capabilities, something Russia quickly made explicit.
During the Cuban missile crisis, President John F. Kennedy was pushed from all sides, including by the joint chiefs of staff, to launch a preemptive bombing campaign against the Soviets to obliterate their nuclear capacity on the island. Instead, JFK opted for a passive blockade and diplomatic back channeling that resolved the issue without kinetic force or the potential for catastrophic escalation. But new AI technology being sold to the U.S. government may forestall a similarly de-scalatory approach in the coming decades.
In June of 2010, General Stanley McChrystal, then America’s top commander in Afghanistan, was forced to resign after a long string of incidents suggested contempt for civilian control of the American military. First, he publicly called for tens of thousands of troops to surge in Afghanistan before a confidential document codifying his opinion leaked to the press, two breaches in protocol that resulted in accusations of subversion of the chain of command. Soon after he and various other commanders and staff were revealed to have disparaged the president andvice presidentv in an article for Rolling Stone. He prepared his resignation the same day it was published.
Now, McChrystal is advising an AI technology firm that bills itself as being better at predicting the future of nuclear warfare than both civilian and military commanders. In a lengthy opinion piece penned in Foreign Policy, McChrystal wrote that the firm Rhombus, where he serves as an advisor, predicted the invasion of Ukraine by Russia better than the American national security apparatus. “The quantity of data we analyze helps predict the next card in an opponent’s deck with previously unimaginable confidence.”
Rhombus CEO Anshu Roy cites Sun Tzu to summarize his firm’s premise: “Battles are won before they start.” Roy’s firm has won a $200 million dollar contract from the U.S. Air Force for “an artificial intelligence platform for strategic decision making in defense and national security enterprises,” and has been sanctioned by the CCP for his firm’s work on behalf of the Taiwanese military. In 2024, Carrier management wrote that Rhombus Power also tracks North Korean missile launches for an unnamed government client, suggesting its potential for left of launch applications has been in motion now for some time.
As The New Republic reported last year, Rhombus’s Ambient product, described as a “digital nervous system, made approximately 32,000 predictions in 2023. Of those tens of thousands, 25,000 were accurate. That’s a .780 batting average, which while great for baseball, might not be high enough for initiating nuclear conflict.
In his article, McChrystal acknowledges that “while AI can make the picture clearer, it only makes decision-makers’ choices more complex”—an apparent nod to civilian control of the military. Nonetheless, there is a subtle un-truth to the idea that AI integration into systems used to detect adversaries’ nuclear command and control infrastructure will be a merely passive tool to provide commanders options in the situation room.
As we’ve seen in the Trump administration’s use of incredibly powerful weaponry and detection technology, intelligence fueled by AI greases the wheels of the national security apparatus, whether the outcomes of kinetic action will have long term beneficial impacts or not. In an interview with Politico, Rhombus officials countered this point by pointing to de-escalating a flare up of hostilities between India and Pakistan by “identifying activity at Pakistani bases that could have been mistaken for nuclear arms escalation…but was not.”
But while the generals advising Rhombus about the promise of AI’s use on the battlefield are charging full steam ahead, not every retired general is in alignment. Prior to his retirement, Lt. Gen. John N.T. Shanahan served as director of the Joint Artificial Intelligence Center, a Department of Defense component studying the integration of AI onto the battlefield.
In an article for the Arms Control Association Shanahan wrote: “One especially destabilizing scenario involves the widespread application of AI to detect and continuously monitor the ground, sea, and airborne nuclear forces of other states. Adversary integration of AI-enhanced systems … could erode confidence in a state’s assured retaliation posture. This loss of confidence could in turn increase incentives for preemptive strikes, weakening the logic of deterrence and undermining strategic stability. The problem becomes even more acute if such surveillance and tracking technologies are paired with AI-enhanced automation of decision-making or launch processes, heightening first-strike incentives during a crisis.”
His conclusion ultimately found that even small scale errors at the input level of AI systems could “produce disproportionate upstream consequences once multiple AI models are embedded across interconnected platforms and decision-support systems.”
As AI integration into “left of launch” ramps up, job listings obtained by The New Republic show the extent to which AI is already being injected into United Strategic Command, which is responsible for maintaining and operating the nuclear arsenal. The command’s official remit includes overseeing Strategic Deterrence, NC3 (Nuclear Command, Control, and Communication) Enterprise Operations, Joint Electromagnetic Spectrum Operations, and Global Strike and Missile Threat Assessment.
These job offers include a General Dynamics listing for an AI engineer to work at Stratcom to reduce the complexity of command and control through AI and machine learning, and an AI advisor to oversee AI integration into the Tomahawk weapon systems threat mission planning center.
Elsewhere, a contractor called DEFCON AI is hiring a subject matter expert to support the development of a modeling and simulation environment who has “deep expertise in U.S. Strategic Command (STRATCOM) programs to help model strategic capabilities, planning constructs, and operational decision frameworks” to “support analysis across the nuclear deterrence mission set.”
Meanwhile, Lockheed Martin posted a listing for a staff AI research engineer based out of STRATcom’s headquarters in Nebraska to work on integrating AI into the nuclear deterrence mission. Taken together, these postings show that AI is being increasingly integrated into both the nuclear command and control infrastructure and into the conventional weapons systems that could be used in a “left of launch” first strike scenario.
AI integration across the Department of Defense has ramped up with executive orders and a new AI-styled space race with great power competitors like China and Russia. But one development over the past two years has integrated AI into the “left of launch” framework with the result of escalating tensions with nuclear powers who claim a new “defensive capability” is an outright escalation on the nuclear battlefield.
In his second term, Donald Trump unveiled his plans for the Golden Dome, a network of space detection satellites, high energy weapons, and ground and sea based interceptor missiles to shield America from missile threats. The program is expected to cost $185 billion, with an initial capability delivered by 2028 and the full system in place in 2035. (Some analysts have put the figure closer to $3.5 trillion.)
The U.S. Space Force has already handed out billions in contracts for prototype interceptor missiles, but the escalatory potential for the system is not found in explosive hardware, but the detection capabilities proposed for the system. Firefly Aerospace and its recently acquired SciTec have touted their products for the Golden Dome, including missile warning systems, surveillance and reconnaissance, autonomous command and control, all in addition to “AI-enabled systems designed for low latency operations to support advanced threat tracking and response across multiple domains.”
Also involved in AI, data processing, and command control systems for the Golden Dome are Palantir, Anduril, and Booz Allen Hamilton. The Space Force General in charge of the Golden Dome, Michael Guetlein, let slip in an interview to the Washington Examiner that “left of launch counterattacks” will also be a core component of the Golden Dome.
Russia and China have taken note, releasing a joint statement criticizing the integration of “left of launch” into Donald Trump’s new project, describing it as a “complete and ultimate rejection to recognize the existence of the inseparable interrelationship between strategic offensive arms and strategic defensive arms, which is one of the central and fundamental principles of maintaining global strategic stability. The project also provides additional impetus to the further development of kinetic and non-kinetic means providing for the left-of-launch defeat of missile weapons and the infrastructure that supports their employment.” In other words, a new arms race is on.
Over the past year, members of congress have introduced tepid legislation to curb the integration of AI into weapons systems and the nuclear arsenal. Senator Kirsten Gillibrand introduced the Secure and Accountable Military AI Act. The bill would essentially mandate something that already exists: a human in the loop for AI targeting and control systems. This will accomplish little more than introduce a new button-pusher into vast and allegedly omniscient AI systems.
Sen. Elissa Slotkin introduced her own bill months before that sought not only to ban AI for the use in a first strike nuclear launch but also to restrict the application of AI to domestic surveillance. But even if these bills were to pass, the full-scale AI arms race with China suggests there may be little hope of stopping the ever-present drift towards a frictionless Department of Defense that moves towards strikes with a will of its own.
As “left of launch” capabilities are streamlined with AI, potential response times for action may decrease. The grace period for escalatory phone calls and meetings also becomes condensed. Both America and its enemies will have access to vast flows of satellite data, social media intelligence, internal communications from hostile actors, and even psychological profiling, all wrapped together into a blaring warning strobe. Maybe this technology really will make us safer, enabling first strike or cyber attacks to prevent and degrade a future launch. But maybe it won’t.
In the Stanley Kubrick film Dr. Strangelove—the most famous on-screen critique of the nuclear arms race—the president asks the mad Dr. Strangelove about the contours of the doomsday machine, an ultimate deterrence mechanism that would bring about the total annihilation of the world. “How is it possible for this thing to be triggered automatically and at the same time impossible to untrigger?” He asks. “Mr. President,” Dr Strangelove replies, “it is not only possible, it is essential.” As AI creeps into “left of launch” operations, in addition to every other nook and cranny of the Department of Defense, the logic underpinning this drive to replace human judgment with a tangle of AI algorithms sounds very Strangelovian indeed.
