Palantir Maven Puts Civilian AI Labs Inside the Kill Chain

Pentagon officials said Palantir’s Maven Smart System now has more than 100,000 users, after it helped U.S. forces strike 13,000 targets in Iran in 38 days. On September 25, 2026, a federal appeals court left Anthropic labeled a supply-chain risk for refusing to drop safety limits on Claude, the model that had already been wired into that targeting stack.

The overlooked parties in this fight are not only generals. They are the software vendor that owns the interface, the model lab that tried to keep two bans, the allies now running the same product, and the officer who still has to click accept when the clock is measured in seconds.

What Maven Smart System Does in a Strike

Maven Smart System is the Pentagon’s flagship AI targeting platform, grown out of Project Maven, which the Defense Department started in 2017 to run computer vision on drone video. Google was an early partner and left in 2018 after staff protests. Palantir took over the software layer and still integrates it. In unpublished company documents reviewed by researchers at the Center for Strategic and International Studies, Palantir wrote that “MSS is powered by the Palantir Platform.”

The interface is a live map. Computer-vision models box objects in satellite and drone video, often inside a yellow outline on the user’s screen. Large language models then sit on top of that layer through Palantir’s Artificial Intelligence Platform, which the firm uses for natural-language queries, document threads, and no-code logic. In 2024, U.S. Central Command’s Maven deployment drew from 179 distinct data sources, a figure industry sources later told CSIS had grown.

Cameron Stanley, the Pentagon’s chief digital and AI officer, described the change in a March 2026 demonstration: operators used to shuffle detections across eight or nine systems. Maven folds find, course of action, and strike tasking into one screen.

We’ve gone from identifying a target to now coming up with a course of action, to now actioning that target. All from one system. This is revolutionary. We were having this done in about eight or nine systems, where humans were literally moving detections left and right.

Cameron Stanley, Chief Digital and AI Officer, March 2026 demonstration

Drew Cukor, the retired Marine colonel who was Project Maven’s founding director, put the old tax in plainer terms. In a March 2026 interview he said crews once spent about 97 percent of their time getting data together and about 3 percent actually analyzing it. A Georgetown CSET review of the 18th Airborne Corps’ Scarlet Dragon exercises found a Maven cell of roughly 20 people matching the output of an Operation Iraqi Freedom targeting cell that had more than 2,000 staff.

MAVEN’S BASELINE TOOLKIT

• Battlespace management: One synchronized map of friendly forces, threats, and civilian activity.
• Target management: A shared list for nominating, ranking, and tracking aim points.
• Deliberate planning: AI-assisted options for timing, weapons, and flight paths.
• Computer-vision detections: Models that label objects in live imagery for a human to confirm.
• Machine-assisted disclosure: Tools meant to speed sharing with partners who need a cleaned picture.
• Generative AI: Language models that summarize intel, draft courses of action, and answer questions in the same window.

Palantir posted Stanley’s walkthrough of that software-as-a-service stack while the Iran campaign was still in its first weeks.

"This is Maven Smart System—Palantir’s software as a service product that we are deploying across the entire department." pic.twitter.com/hIaQAiq4iJ

— Palantir (@PalantirTech) March 12, 2026

The demo is the product pitch. It is also the thing a targeting miss has to survive, because the same clicks that compress hours into seconds are the clicks that send a weapon.

Claude Reached Classified Networks Before the Ban

On November 7, 2024, Palantir, Anthropic, and Amazon Web Services said they would bring Claude models to classified networks inside Palantir AIP on an Impact Level 6 environment. The release named Claude 3 and 3.5. Shyam Sankar, Palantir’s chief technology officer, said the company was the first industry partner to put Claude into those rooms, and the copy stressed that officials would keep “their decision-making authorities.”

CSIS later wrote that Anthropic was the first frontier lab to deploy models in those classified settings and had powered Maven’s LLM functions into 2026. An anonymous National Geospatial-Intelligence Agency official, quoted in Katrina Manson’s reporting and cited by CSIS, said language models added a fivefold jump in targeting speed on top of the tenfold gain computer vision had already delivered. In July 2025 the Pentagon signed a $200 million, two-year prototype deal with Anthropic for frontier AI work.

Then the customer and the lab stopped agreeing on what “lawful use” meant. Anthropic CEO Dario Amodei told Defense Secretary Pete Hegseth he would not let Claude be used for fully autonomous weapons or for mass surveillance of people in the United States. The department wanted the models for every lawful purpose. On February 24, 2026 it set a February 27 deadline. Anthropic did not yield.

HOW THE LAB LOST THE VETO

1. November 7, 2024: Palantir, Anthropic, and AWS announce Claude 3 and 3.5 inside classified AIP.
2. July 2025: The Pentagon signs a $200 million, two-year Anthropic prototype agreement.
3. February 24, 2026: The department demands unrestricted lawful use by February 27.
4. February 27, 2026: Hegseth directs that Anthropic be treated as a supply-chain risk, with a wind-down window for existing work.
5. February 28, 2026: Operation Epic Fury begins. Claude is already inside the Maven workflows used for the opening strikes.
6. March 3, 2026: The designation is announced as effective immediately for military-related contractor work.
7. September 25, 2026: A 2-1 D.C. Circuit panel upholds the Pentagon’s conclusion that keeping Claude in its systems presented a national-security risk.

Judge Gregory Katsas, joined by Judge Neomi Rao, wrote that the department had “ample support” for that finding. The risk the court accepted was not a foreign back door. It was Anthropic’s remaining power to block uses it had written into contract, including the two bans Amodei would not drop. Defense Undersecretary Emil Michael put the official line more bluntly after the ruling, saying warfighters would sleep better knowing “no private company will insert their opinions in the chain of command.”

Hegseth had already framed the lab as a hostage-taker. “America’s warfighters will never be held hostage by the ideological whims of Big Tech,” he wrote when the designation went public. A label long associated with firms tied to adversaries was now being used on a U.S. model maker because its safety policy could, in the department’s telling, shut a tool off or hem it in. Anthropic said it was considering further review. Palantir, the company that owns the screen, was not in the dock.

A Thousand Day-One Targets Left 86 Seconds

CSIS, drawing on campaign reporting, said the United States used Maven to help strike more than 1,000 targets in the first 24 hours of the Iran war, a tenfold jump over the pre-Maven era. One thousand aim points in a day is about 86 seconds each if you divide the clock evenly, which no real targeting cell does. Stanley said Maven then helped the force hit 13,000 targets in 38 days of Operation Epic Fury, about 342 a day across that whole stretch, which is a different tempo from the opening surge.

A Pentagon spokesperson said unclassified Maven use rose 38 percent during the campaign and classified use rose 89 percent, with peak daily token burn of about 20 billion, up 4,425 percent from the prior baseline. James Mazol, deputy undersecretary of defense for research and engineering, told the DefenseTalks conference on September 22, 2026 that users had gone from about 50,000 in January to over 100,000 after Epic Fury kicked off. “Everybody wants this because once they see the power of it, they want to sign up as quickly as they can,” he said.

MAVEN AT WARTIME SCALE

Deputy Defense Secretary Steve Feinberg’s March 2026 memo told the department to turn Maven into a program of record by the end of fiscal 2026 and to move oversight from the National Geospatial-Intelligence Agency to Stanley’s office. The original May 2024 award was a $480 million, five-year Palantir contract. The ceiling move in May 2025 is what locked in nearly $1.3 billion through 2029. At that point the model lab is a drop-down. The integrator is the program.

Why the Minab School Was on the List

On February 28, 2026, the first day of the air war, missiles hit Shajareh Tayyebeh School in Minab, near the Strait of Hormuz. Forensic Architecture reconstructed the Minab school site from photos, video, and satellite imagery and found that more than 121 children, aged 6 to 13, were killed along with 26 teachers. The group showed the building had been converted into a school by October 2015, with playgrounds and blue and pink compound walls visible in later years, and that nearby lots were in civilian use, including a clinic and a cultural centre.

Witnesses collected with Sky News put the first wave around 11:22 a.m. The first impact landed about 130 meters from the school; a later round in that same short sequence hit the school itself. A second wave landed around 2:30 p.m., after Iranian channels had already described the site as a girls’ school. Bellingcat identified an RGM/UGM-109 Tomahawk in video of that later round, and munitions specialists working with Airwars and Armament Research Services matched fragments to the same family. The United States was the only party in the war known to field that missile.

WHAT WE KNOW

• The building: Open imagery shows a working school for about a decade before the strike, not a newly marked shed.
• The weapon: Independent munitions analysis identified a U.S. Tomahawk in the Minab sequence.
• The toll on site: Forensic Architecture counted more than 121 children aged 6 to 13, and 26 teachers, among the dead.

WHAT IS UNCONFIRMED

• Maven’s exact click: Campaign reviews have said outdated files still listed the site as an Islamic Revolutionary Guard Corps facility and that Maven helped compress day-one lists from hours to minutes, but a public artifact showing the school as a machine-nominated aim point has not been released.
• Who signed: Names of the officers who accepted that aim point, and whether a civilian-harm cell saw it, remain inside the command.

The hole is the point for the vendors. If the miss was stale data plus a human in a hurry, the software still set the pace of that hurry. If Maven surfaced the file among opening-day recommendations, then the model and the map did more than fetch video. Either way, the people with legal exposure are the ones who accepted the nomination, not the lab that had already been told it could not keep a veto.

In late September, seven months after Minab, Central Command changed how it refreshes planned targets. The package described to the press included new open-source feeds meant to track civilian movement and Maven upgrades designed to re-check intelligence for facts that should disqualify a site. The fix, in other words, was more of the same platform, pointed inward.

Palantir Holds the Contract That Stays

That is the hidden stake. Anthropic’s product was the reasoning layer that made unstructured reports and course-of-action chat possible at war speed. Palantir’s product is the room those models live in. When the lab and the Pentagon split, the department could designate the lab a risk. It could not, at 100,000 users, rip out the map.

Palantir sells that stickiness as virtue. Its defense AIP pages promise an audit trail for human-machine teaming, plus controls over what a model is allowed to see. Sankar’s 2024 line about keeping decision rights with officials is the same claim in softer clothes. The Iran war is the first large test of whether those rights are real when the queue is a thousand targets and the interface is built to make “accept” the easy key.

Our partnership with Anthropic and AWS provides U.S. defense and intelligence communities the tool chain they need to harness and deploy AI models securely, bringing the next generation of decision advantage to their most critical missions.

Shyam Sankar, Chief Technology Officer, Palantir, November 7, 2024 partnership release

A tool chain can be swapped on a slide. It is much harder to swap after every combatant command, the National Guard Bureau, and a wartime user base that doubled in a few months have already learned one screen. Mazol’s September remarks were about proliferation and compute, not about pausing for a new legal theory of “should we.” Classified compute, he said, is now the bottleneck, because the department cannot train on secret data in ordinary commercial halls.

Replies to Palantir’s own March demo went immediately to the school. That is not a sideshow. The product’s public face is a fused battlespace that turns detections into weapons advice. The first-day miss in Minab is what that face looks like when the underlying file is wrong and the clock is not built for a long argument.

The Alliance Now Runs Maven at SHAPE

On March 25, 2025, the NATO Communications and Information Agency finalized a deal so Allied Command Operations could run a variant called Maven Smart System NATO. The agency announced it on April 14, 2025, and said the buy took six months from requirement to contract, among the fastest in NATO’s history. Ludwig Decamps, NCIA’s general manager, said the point was to put customized AI in the hands of allied forces. Palantir counselor Shon Manasco called it an AI-enabled warfighting platform at Supreme Headquarters Allied Powers Europe.

NCIA said the system would use a range of AI, including large language models, for intelligence fusion, targeting, battlespace awareness, and faster decisions, and that ACO should be on it within about 30 days of the announcement. In March 2026, Vice Adm. James Morley, deputy commander of Joint Force Command Norfolk, said Maven was already in at SHAPE and at Allied Joint Force Command Brunssum, with Norfolk due to take it in the spring. The alliance did not wait for the Anthropic court fight to finish. It bought the Palantir layer.

That is how a U.S. vendor dispute becomes a coalition problem. Allies already manage different legal tests and no-strike lists with caveats and red cards at human speed. A shared Maven instance asks them to trust not only an output but the way the list is built, ranked, and pushed to a shooter. If one capital will not accept a model that can be throttled by a San Francisco use policy, and another will not accept a model with no such throttle, the argument is no longer technical. NATO acquired Maven Smart System NATO as common warfighting software. Common software is common doctrine, whether parliaments voted on it or not.

NCIA is pleased to team up with Supreme Headquarters Allied Powers Europe (SHAPE) and Palantir to deliver MSS NATO to the Warfighter, providing customized state-of-the-art AI capabilities to the Alliance, and empowering our forces with the tools required on the modern battlefield to operate effectively and decisively.

Ludwig Decamps, General Manager, NATO Communications and Information Agency, April 14, 2025

The Declaration Still Wants Human Command

The paper trail still talks as if a person is in charge. The U.S. State Department’s Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy, launched at the 2023 REAIM meeting in The Hague, sets ten measures for responsible military AI. Endorsing states say military AI must sit in a responsible human chain of command, stay auditable to defense staff, and be used by people trained to know the limits of the system so they do not rubber-stamp it. High-consequence tools are supposed to get senior oversight. Safety features on systems that keep learning are supposed to be watched so they do not quietly decay.

Those lines are not a statute. They also do not name Palantir, Claude, or a 86-second queue. They do describe the fight the Pentagon just won in court in reverse. Anthropic’s position was that some uses stay off limits even when a customer calls them lawful. The department’s position, now blessed by a split appeals panel, is that a vendor who can enforce that kind of limit is itself the risk. The declaration’s language about a human chain of command does not say which human, and it does not say what to do when the human’s job is to clear a list the machine already ranked.

Palantir keeps the program of record, Anthropic’s two banned uses are now a court record, and the officer who accepts a nomination still owns the strike.