Meet the Gods of AI Warfare

The rise of AI warfare speaks to the biggest moral and practical question there is: Who—or what—gets to decide to take a human life? And who bears that cost? In 2018, more than 3,000 Google workers protested the company’s involvement in “the business of war” after finding out the company was part of Project Maven, then a nascent Pentagon effort to use computer vision to rifle through copious video footage taken in America’s overseas drone wars. They feared Project Maven’s AI could one day be used for lethal targeting. In my yearslong effort to uncover the full story of Project Maven for my book, Project Maven: A Marine Colonel, His Team, and the Dawn of AI Warfare, I learned that is exactly what happened, and that the undertaking was just as controversial inside the Pentagon. But that didn’t slow its forward march. Today, the tool known as Maven Smart System is being used in US operations against Iran. How the US military’s top brass moved from skepticism about the use of AI in war to true believers has a lot to do with a Marine colonel named Drew Cukor. In early September 2024, during the cocktail hour at a private retreat for tech investors and defense leaders, Vice Admiral Frank “Trey” Whitworth found his way to Drew Cukor. Now Project Maven’s founding leader and his skeptical successor were standing face-to-face. Three years earlier, Whitworth had been the Pentagon’s top military official for intelligence, advising the chairman of the Joint Chiefs of Staff and running one of the most sensitive and potentially lethal parts of any military process: targeting. Colonel Cukor, an intense Marine intelligence officer described to me by one of his seniors as “a one-man wrecking ball” who took on military orthodoxy, defense bureaucracy and the pursuit of AI warfare to his own cost, was wrapping up his five years as Project Maven’s chief. In a meeting so tense some present had squirmed, I learned that Whitworth—an exacting former SEAL Team 6 intelligence director who sat on the military targeting committee for nearly two decades—had drilled Cukor about whether Maven and its use of AI was skipping crucial steps in the targeting process, moving too fast and bending rules. “Tell me about what happens after the bad drop when we go through a congressional [hearing] and we’re getting hard questions?” Whitworth demanded. He worried about record-keeping and accountability when it came to involving AI in targeting, and he expressed strong doubt that Project Maven was worth the billion dollars Congress had already spent on it, much of which had gone to Silicon Valley’s controversial upstart darling: Palantir. When Whitworth took charge of the National Geospatial-Intelligence Agency in June 2022 and became responsible for the future of Project Maven following Cukor’s departure, he still worried that Project Maven was overpriced, overhyped, and incautious about the targeting principles he most cared about. Whitworth could have shut the program down in a heartbeat. The future of Cukor’s baby looked increasingly in doubt. “We were all very concerned,” Cukor told me in one of our weekly afternoon talks over the course of more than a year. “Trey was not a friend.” I would come to see Cukor as a leading historical figure in a war that hasn’t happened yet. That seemed to be what almost everyone to do with Project Maven thought, whether they feted or hated him. Alex Karp, Palantir’s chief executive, referred approvingly to Cukor as “crazy Cukor” and called him “the founding father of AI targeting.” After his showdown with Whitworth, Cukor told others: “I will either be famous or live in infamy.” But now, more than two years into leading the NGA and more than two years into Russia’s war against Ukraine, rather than abandoning Maven, Whitworth praised the program. “Drew, this is important work,” he assured Cukor at the September 2024 event. Maven Smart System—the software platform built by Palantir that brought together disparate battlefield and other data on a digital map and displayed AI detections that could be deployed in targeting—was adaptable. It could integrate with any system and become new with each software update. It could do what people wanted. Cukor described the vice admiral as methodical, saying that Whitworth had just reasoned his way to endorsing Maven. Cukor thought Whitworth had come to understand why the US needed to bring AI into the targeting cycle. (The portion of Maven’s $250 million annual budget that went to NGA may also have helped, he thought.) “It speaks to his character, honestly,” said Cukor. “It wasn’t an apology as much as a formal recognition. We didn’t hug, but it was an important conversation.” Under Whitworth, Maven would have its coming-out party, emerging from years of secrecy tightly maintained under Cukor in the wake of the Google protests. Six months earlier, deciding to shoot comprised the shortest element of the targeting cycle. Now every other part of the cycle was so close to being automated, and so compressed in time, that deciding to shoot was the lengthiest part. Internal documentation referred to Maven ATR: automatic target recognition. In public, Whitworth started describing Maven as his agency’s “marquee targeting program of record.” A few days after he spoke with Cukor, Whitworth stepped onto a stage for a livestreamed Palantir customer event. He could hardly have cut a stronger contrast with the Palo Alto crowd: His service dress blues came with gold buttons, gold threads round his sleeves, and bright ribbons. In his high-shine formal black shoes, he stood in front of a cabinet displaying colorful Nike sneakers. His talk on Maven Smart System followed directly after two other Palantir customers, one who leased railcars and another who supplied automotive seating. War was now just another business process, sandwiched between sales and health care. Amit Kukreja, a prominent Palantir commentator, investor, and fan of the company’s “merch,” was narrating the event live off to the side. He described it as a “new and special” moment for Palantir’s retail investors to learn about the company’s government work. Even Karp appeared taken aback. “I didn’t even know we were allowed to talk about this stuff,” he said, after laying claim to the “most elite and interesting” government clients in the world. Palantir had already won an Army contract with a $480 million ceiling for Maven Smart System that spring, and later would win another to supply the system to all military services in September for up to $100 million. In spring 2025, the Pentagon’s contract ceiling for Maven Smart System was raised to $1.3 billion, due to run until 2029. And NATO said it would become a customer for Maven Smart System too. Ten NATO customers were thinking about buying the system for their own country. The UK would reportedly sign a £750 million (roughly $1 billion) deal for Palantir’s military AI tools during a high-profile state visit from Donald Trump in September 2025. Up on the Palantir stage, Whitworth talked through AI targeting as a screen beside him played a demonstration. An icon flashed up alerting the audience to “Possible Enemy Activity.” A cursor click revealed a group of tanks over a “notional” demonstration map of Kherson in Ukraine. The tanks were four clicks from evisceration. Palantir’s Target Workbench popped up. Two more clicks established the tank group’s height, latitude, and longitude; and then paired the target with an “effector” (in this case an F-22A fighter jet 82 miles away). One more click and a green tick flashed up: “Target destroyed.” Nearly a year later, on a hot day in the high summer of 2025, I stepped into NGA’s headquarters at the Fort Belvoir Army Base in northern Virginia. It was my second visit to the spy agency HQ, and I wanted to find out why Whitworth had changed his mind, how much Maven had spread, and how Maven’s new backers saw the risks and rewards of mainstreaming AI into military workflows. By then, Whitworth had become so ardent a fan of AI that his agency was pumping out machine-produced intelligence reports for US decisionmakers that “no human hands” had touched. And the NGA had launched a $708 million contract for data labeling in support of Maven’s computer vision models, the largest such appeal in US history, that would ultimately go not to self-made billionaire Alexandr Wang’s Scale AI but to Enabled Intelligence, a startup focused on hiring people on the autism spectrum expert in pattern recognition and comfortable with repetitive work. My visit required the rigmarole of any meeting at a spy agency. Courteous background checks and vetting; no phone, laptop, or smartwatch allowed; and one step more curious: writing down not only the make and model, but also the serial number inscribed on my tape recorder, which I resolved never to use again for any interview after the visit. The building was a temple to geospatial intelligence, or GEOINT, the pursuit of insightful analysis tied to locations on a map. A mesh of reflective glass encased by nearly 2,000 concrete triangles covered the blast-resistant facade, as if each one were attempting to triangulate a different location. More than 8,500 personnel worked at headquarters, but I was there to meet four particular NGA officials. Each, in their own way, was deeply involved in the development, standards, and spread of Maven. It was, I was told, unprecedented for them to gather all in one room to brief a journalist on Maven, and I was eager to hear what was at stake for them. “This is our reputation on the line,” Whitworth told me in the interview. After he saw how easy it was to integrate the system into combat scenarios, it didn’t take long for him to change his mind: “I started really believing in it.” Far from being sheepish about ushering in a new age of AI warfare, its midwives wanted their names stamped over it. Some had become quite “ornery” in pursuit of credit, one NGA official said. I wondered if NGA wanted its fair share, mindful that some advising the second Trump administration wanted to wrest control of Maven and AI away from NGA and back into the Pentagon. “There’s no one person who can claim credit for this thing. It’s too big.” The NGA officials walked me through Maven’s developments since the agency took over most of it two years before. Five of eight Maven initiatives, including analyzing drone feeds and satellite imagery, ended up with NGA. Whitworth wanted to expand the scope and capabilities of his agency in line with the expansion of ubiquitous global sensors. AI relied on data, and that required global surveillance to deliver it. While NSA could listen in to the world, NGA could watch it. Whitworth made clear he wanted to do that in minute, constant detail—surveilling the entire globe, at all times. NGA previously gave me a demonstration showing how AI could flag military construction in China—such as the arrival of a new rail depot at a missile base. NGA kept track of all movement at 49,000 airfields around the world. Whitworth even wanted to put GPS, or a similar navigation system, on the moon. And if GPS got jammed or hacked, he wanted other ways to map space too: NGA was fashioning digital maps drawing on magnetics, gravity, remote sensing, celestial navigation, and elevation. “From seabed to space,” went the new mantra he unveiled in 2023. The US war horse wanted omniscience, omnipresence, and omnipotence. Nearly two years into its work in support of Ukraine, Maven became a “program of record” at the beginning of November 2023. That was Pentagon-speak for a fully funded line of effort with backing from Congress. It came with the expectation of a consistent budget for the coming years. The lines were still blurred: The Pentagon’s successor to the JAIC, the Chief Digital and Artificial Intelligence Office, paid the licenses for Maven Smart System and managed the text-based parts of Maven such as “reading” captured enemy material, while NGA produced the computer vision models that showed up on Maven Smart System screens. Throughout 2024, Whitworth drummed up new users for the platform. He rang combatant commanders in every region to tell them what NGA was adding, shopping Maven’s latest features. Addressing criticism I’d come across, he insisted that Maven was useful not only in Europe, but also in the Indo-Pacific, and for moving targets as well as static targets. Maven Smart System had particularly taken off in the Middle East. General Erik Kurilla had started using the platform “extensively” in support of US weapon strikes, once he took over US Central Command in April 2022. He hired former Google AI expert Andrew Moore, and spent much of 2023 practicing how to get through a thousand targets a day, cooperating with the UK and others in a series of experimental 90-day sprints. In early 2024, I learned that the command had made “a pretty seamless shift” from experimenting with the platform in exercises to doing all this in combat. Here would be the first real US test of AI at war on a large scale. “October 7th everything changed,” Schuyler Moore, the chief technical officer at CENTCOM, told me, referring to the deadly 2023 Hamas attack on Israel that international rights groups said constituted war crimes and crimes against humanity. “We immediately shifted into high gear and a much higher operational tempo than we had previously,” she said. Brigadier General John Cogbill, CENTCOM’s deputy director of operations—who served under Kurilla, including at the 75th Ranger Regiment—put it this way: “It’s just been off to the races ever since.” In February 2024, the command used Maven Smart System to locate rocket launchers in Yemen and unmanned surface vessels in the Red Sea. Moore told me Maven’s AI helped narrow down more than 85 targets that US bombers and fighter aircraft subsequently struck in Iraq and Syria, in reprisal for the death of three US service members in Jordan the month before. It was public confirmation that the US military was using AI to identify enemy systems for its own weapons to strike. “We were using these tools in a way that we’d never used them,” Cogbill told a podcast the following month, saying the command became “hyper-focused” on Israel. By 2024, the command had 179 different live data feeds from land, sea, air, space, and cyber pouring into Maven Smart System. CENTCOM is using it most, Whitworth told me. That region alone had 13,000 accounts, with 2,500 people counting as regular users who log in “at least a few times a week,” said Rear Admiral Liam Hulin, a former commander of SEAL Team 3, who was now deputy director of operations at CENTCOM. Maven could also discern the nearest available weapons, the most suitable ones for the task, flying time, weapons loading details, and the whereabouts of personnel and partners. Operators would click through Maven’s Target Workbench, approve or disapprove targets, sequence them according to priorities, and send a message direc