The Pentagon Asked Palantir for a Kill Chain. Maven's Founding Director Agreed, On the Record.

Drew Cukor is a retired Marine Corps colonel and the founding director of Project Maven, appointed under the Algorithmic Warfare Cross-Functional Team. He is not an NGA official. NGA is the GEOINT customer and inherited the programme's GEOINT components in January 2023, and the only NGA official quoted on the record in the public material I found is a different, unnamed person, cited by CSIS for the fivefold speed figure. Conflating the two would have made this post look better and would have been wrong.

Here is the sourcing rule I have applied throughout. What Manson documents, I attribute to Manson. What the procurement record documents, I attribute to the record. Where I could not verify something I have cut it, including a pull quote I had wanted badly enough to keep.

What the founder said, and what he did not say#

The claim I need to test is that Project Maven did not merely execute targeting policy faster, it made policy options available that would not otherwise exist, and nobody with authority reviewed that change.

Cukor's on-record position across several 2026 interviews is that this is exactly what the programme was for. CSIS identifies him as "the retired Marine colonel who served as the founding director of Project Maven," and The Verge described him in April 2026 as "a Marine intelligence officer." He is a well-documented figure with a documented view, which is a stronger foundation than the vague "former program director" framing I started this post with.

Now the part that complicates my own thesis, and it comes from a source I am otherwise leaning on. Matthew Manson, the former chief digital and artificial intelligence officer, has given a materially different accounting of human involvement. In a TechPolicy.Press interview he described humans present at six points in the process, with computers replacing humans at three of those places, and a human becoming supervisory in one. That is not the same as "Maven performs the first four of six kill chain steps." The popular formulation flattens a mixed picture into a clean fraction, and Manson's version, which leaves three points machine-run, is the more alarming one if you care about who is accountable when the machine is the actor. Democracy Now carries a substantive interview on the same ground.

So the state of the evidence is two accounts from people with standing, and they disagree about the shape. Nobody published a reconciliation, and nobody asked for one.

What the agency running it says it is#

NGA's own GEOINT AI page, the participating agency's public self-description, is carefully bland. In full:

"NGA Maven was established in 2017 as the Pentagon's flagship AI project to integrate AI into military workflows. The GEOINT aspects of Maven were entrusted to NGA in January 2023. Its state-of-the-art computer vision and AI capabilities are now integrated into various military analytic workflows to automatically detect, identify, characterize, extract, and attribute features and objects in imagery and video. NGA Maven is already producing large volumes of computer vision detections for warfighter requirements across multiple operational locations. It's generated millions of data labels and lowered latency detections."

Note "warfighter": Maven's output is not intelligence for analysts but detections delivered to operational users. And "attributed," not "identified," is a deliberate term of art: attribution is an inference about who or what a detected object is, the step where machine error becomes human harm.

The volume problem is real and structural. NGA's GEOINT data intake is near tripling, which is why ASPEN, the Analytic Services Production Environment for the NSG, exists: a program of record running a five-year modernization of intelligence production begun in May 2023. Maven and ASPEN are the agency's answer, and I will not write this as a cynical-only story because the operational justification is genuine and measurable. But the answer's scope was never re-examined as it grew, from annotating imagery, to four of six kill chain steps, to domestic repurposing for border control and drug policing. The trajectory, not the justification, is what needs explaining. Nobody explained it.

The procurement record, which anyone can check#

The spine of the piece; every figure traces to a public solicitation or award record.

I pulled this list myself from NGA's contract announcements page, sam.gov and USAspending.gov, and I have cut four rows I could not verify to the standard the rest of the table meets. A $708.3M Sequoia data-labeling IDIQ, a $646M Glacier Bay award, a UK Ministry of Defence strategic partnership, and a $13 billion self-directing systems budget request all appear in secondary coverage, and I could not tie any of them to a dated award record. They may well be real. I am not publishing them in a table whose entire function is to be checkable.

The initial $480 million five-year IDIQ went to Palantir in May 2024, Army Aberdeen Proving Ground as awarding agency, Office of the Secretary of Defense as funding agency. Initial coverage: five combatant commands, CENTCOM, EUCOM, INDOPACOM, NORTHCOM/NORAD and TRANSCOM, plus the Global Information Dominance Experiments.

DefenseScoop, 23 May 2025: Pentagon leaders raised the ceiling by $795 million "to prepare for what they expect will be a significant influx in demand from military users for the AI-powered software capabilities over the next four years." A defense official: "In response to this growing demand, the CDAO and Army increased capacity to support emerging combatant command operations and other DOD component needs." The ceiling is recorded as $1,275,000,000 running through 28 May 2029.

One discrepancy: SpaceNews, DefenseScoop and CSIS quote the ceiling as $1.3B; other reporting says $1.7B. I go with $1.3B, since that is the figure the award record carries; the resolution belongs to sam.gov and USAspending.gov.

NGA Director Vice Adm. Frank Whitworth, at the 2025 GEOINT Symposium per SpaceNews on 22 May 2025, stated there were more than 20,000 active Maven users across more than 35 military service and combatant command software tools in three security domains, and that the user base had quadrupled since January. On posture: "While we've treated AI with the care and power it deserves, we've entered into an 'unchained phase' by implementing a whole-of-agency approach to AI transformation."

"Unchained phase." From the director of the agency whose data is being processed.

Four of six, stated carefully#

The kill chain has six steps: identify, locate, filter down to lawful valid targets, prioritise, assign them to firing units, and fire. In 2024, reporting said Maven can perform four of them, and CSIS repeats that framing verbatim. Matthew Manson's own accounting is different in a way that does not reduce the concern, as set out above.

Start with the limit: the two steps Maven does not perform are the last two. Human authorisation and weapon release remain with people. The Pentagon says humans are in the loop, and on the evidence that appears literally accurate. If this post hid that, it would be a Rube Goldberg of accusation.

What actually changed is throughput, not clicks#

From The New Yorker's review of Manson's book: a "single click" could send coordinates through a tactical data link to a specific weapons platform so that it could fire at the target. The entire process, from target identification to target destruction, is four clicks. In 2023, one source told Manson he could sign off on eighty targets in an hour: "Accept. Accept. Accept." The old system could hit fewer than a hundred targets a day.

The Independent reported on 10 March 2026 that Maven Smart System "is currently able to make 1,000 targeting recommendations in an hour." Recommendations, not strikes. That is still the figure to sit with, because a recommendation that a human never accepts changes nothing operationally and a human who accepts a thousand an hour is not reviewing anything. Army leadership told Bloomberg in May 2025 that it wanted Maven to help troops make 1,000 battlefield calls in a single hour, so the demand-side target and the reported capability are the same number.

CSIS, in "What Is Maven Smart System, and What Does It Do?" on 6 February 2026: "During the first 24 hours of the war in Iran, the United States used Maven Smart System (MSS) to help strike more than 1,000 targets, a tenfold increase over what was possible in the pre-MSS era." The same piece, citing Manson's book, relays an anonymous NGA official saying such capabilities produce a fivefold increase in targeting speed, adding to the tenfold increase computer vision already provides.

The technical heart of the post: CSIS identifies what LLMs add to MSS: "natural-language interaction and broader pattern analysis." Computer vision's tenfold increase is a better camera. Language models let an operator query the system in plain language and get back patterns no one specified in advance, and interrogating it at conversational speed is a different instrument from a faster camera. The governance framework was written for the first instrument; nobody revised it when the second one shipped.

The consequence nobody quantified#

The throughput numbers are the part that matters, because on their own they are defensible: machines detect faster, humans authorise, the system is faster at the boring part. What is not defensible is the absence of any published decision about what rate is acceptable as a policy input. Nobody has written down the number at which a human sign-off stops being a judgement and becomes a rubber stamp, and the reported capability has already moved to a thousand recommendations an hour.

Richard Baker, the defense reporter who has covered this for years, in the same New Yorker piece:

"The real question, the question almost nobody was asking, is not about Claude or any language model. It is a bureaucratic question about what happened to the kill chain, and the answer is Palantir."

If throughput is a precondition for policy, the right question is not "is the human in the loop?" It is in the loop. The right question is: who decided that a thousand recommendations an hour was an acceptable policy input, and what process was supposed to review that decision?

Nobody did. That is the finding.

Google's refusal, and why it did not matter#

In 2018, Google employees staged a massive walkout over its work on a primitive iteration of the project. Google did not renew the contract; Palantir took over. And following its work on Project Maven, Google introduced AI principles stating it would not design or deploy AI for weapons or technologies intended to cause injury.

Now read the sequence:

1. A capability is proposed for analyst assistance: helping people find things in imagery.
2. A major technology company refuses on stated ethical grounds, and the refusal becomes a written corporate principle.
3. The program continues and grows for eight years.
4. It accumulates a $751.5M contract, then a $480M contract, then a $1.3B ceiling.
5. It becomes a program of record: a formal, budgeted Pentagon commitment.
6. NATO adopts it. Ten member nations pursue it.
7. It is repurposed domestically for border control and drug policing.

Nothing in that sequence required a decision, which is exactly the problem. The institutions that could have said no, Google's engineers, the CDAO, the Pentagon, the Congress, each had grounds and none acted. The system did not need authorisation. It needed a budget, and the budget arrived.

The 2018 protest did not fail for lack of an argument. It failed because the program continued under a different vendor. The objectors were objecting to the capability, and the capability was indifferent to who built it. That generalises past Maven: vendor ethics is a lever that does not work on capabilities. Google could refuse; it could only stop Google.

The scope expansion, which is the part that should scare you#

From CSIS, 6 February 2026: NATO now has its own Maven contract with Palantir, supporting Allied Command Operations, "and that prompted ten member nations to pursue one." Per Manson, Maven is being repurposed for border control and drug policing at home. Those are the two expansions I can trace to a dated source; a UK Ministry of Defence strategic partnership and a Pentagon self-directing systems budget line have also been reported, and I have left them out because I could not verify either to the standard of the rest of this post.

The normalisation is faster than the governance, and it held at every step of this program's history.

The Anthropic episode, which is the failure that is actually documented#

Palantir announced a partnership with Anthropic in November 2024 to bring Claude to classified networks through AIP. The Pentagon renewed and expanded that work in a July 2025 contract. Reporting indicates Claude was used through MSS in the 2026 Iran war.

Then it gets worse, and here I have to be careful with the sourcing, because the sentence that circulated is garbled and the direction of the interpretation matters. When The New Yorker asked the administration about Maven's role, one administration official is reported to have read it as a signal that Anthropic, then renegotiating its own federal contract, "was perhaps a faithless partner." That is an official's interpretation of an event. Nobody has established who interpreted what, and I am not going to build the argument on the confusion.

A Pentagon program-of-record vendor's commercial conflict with a model provider, mid-war, was read by the administration as a loyalty test. That is what happens when a critical infrastructure dependency is a single company, and the company's other customer is a different government.

Congressional Democrats sent a letter to Hegseth demanding a detailed account of how A.I. was used in the Iran campaign. Whether that accounting ever came is something a functioning press corps should retrieve in an afternoon; the difficulty of retrieving it is itself the answer.

The structural risk is not the sci-fi one. The intelligence community's targeting layer now depends on a commercial platform supplied by one vendor, integrated with a frontier model supplied by a company that also sells to adversaries of the United States, and the governance of that arrangement was never written down because no one was ever required to write it. Not the model. The procurement chain. The chain itself.

Where this could be wrong#

"The human is in the loop, so this is all alarmist." The best objection first. Maven reportedly performs the first four of six steps; weapon release and target assignment remain human, and the Pentagon's claim appears literally true. But the finding is not about whether a human signs off. It is about whether a thousand sign-offs an hour is a policy input anyone reviewed.

"I am accusing a serving officer of war crimes." No. Cukor is retired, and his position is about throughput and policy formation, not unlawful conduct. I am not presenting it as one.

"Civilian casualties are not established for any of this." Correct, and the most serious risk here. A school attack in the 2026 Iran war is attributed in some reporting to operations in which Maven was involved, and I deliberately decline to build on it: casualty attribution in an active war involving AI-assisted targeting is contested and requires evidence this post does not have. Getting a casualty attribution wrong loses the reader permanently; the analytical contribution survives without it.

"Every one of these numbers came from a book by an investigative reporter." Fair, and this is why I split sourcing: Manson establishes the operational detail; the public record, NGA's announcements, sam.gov, USAspending.gov, establishes the scale, the ceilings, the dates, and the 20,000-user figure.

"Programmes of record are normal acquisition milestones." They are. But the step immediately before the designation, the decision that it should exist in this form, has no owner in the documented record. The designation formalises an outcome whose evaluation was never performed.

"Every military does this. This is normal capability development." Largely true for any single system, and I concede it. The finding is about conjunction and rate: normalisation across allies, domestic repurposing, a four-of-six kill chain role, a fivefold LLM speed multiplier on top of a tenfold computer vision gain, a reported thousand recommendations an hour, and no located process for evaluating any of it. The bundle is the story.

A note on what I cut, because the cuts are part of the finding. Four rows of the procurement table, a Hegseth letter I could not retrieve, a Feinberg memo I could not date, a purported Maduro capture, and a pull quote I wanted badly enough to keep that I could not locate in the source it was attributed to. What survives is thinner than what I started with and more defensible.

The kill chain lost two steps, and the question of whether that was acceptable was never asked by anyone with the authority to ask it.