The System That Wars Are Fought With Now
In the spring of 2026, as U.S. and Israeli forces conducted simultaneous strikes across the Iranian theater, a software platform running on classified Pentagon cloud infrastructure was doing something extraordinary: it was telling warfighters where to aim. Not alone — there are human operators in every loop, at least officially — but with a speed and confidence that, by all accounts, has fundamentally restructured how America wages war.
That platform is Project Maven, and its primary contractor is Palantir Technologies. According to reporting from Bloomberg, the Washington Post, and Democracy Now, the Trump administration has struck more than 11,000 targets in Iran since the conflict began. That volume is not achievable with traditional, manual intelligence-to-strike processes. It requires automation — and Maven provides it.
For the first time in a major peer-adjacent conflict, AI is not just supporting the fight. It is architecting it. Understanding what Maven is, how it actually works, and where it is failing matters enormously — not just for military analysts, but for anyone who cares about the future of warfare and the laws governing it.
What Is Project Maven, Really?
Project Maven — formally the Algorithmic Warfare Cross-Functional Team (AWCFT) — was born in 2017, launched by then-Deputy Secretary of Defense Bob Work as a crash program to bring machine learning into intelligence, surveillance, and reconnaissance (ISR) workflows. The initial impetus was simple and strategic: the United States was drowning in drone footage. Analysts tasked with watching feeds from MQ-9 Reapers and Global Hawks were burning out, unable to manually review the terabytes of video being collected every day. Maven's first job was to help computers watch video so humans didn't have to.
Early versions of Maven used Google's TensorFlow platform to build object-detection models that could flag vehicles, personnel formations, weapons systems, and structures of military interest in drone video. Google's 2018 withdrawal from the contract — triggered by employee protests — temporarily disrupted the program, but the Pentagon quickly moved Maven's technical development to other partners, including Microsoft, Amazon Web Services, and ultimately Palantir, which won a major contract to take the system to its next generation: the Maven Smart System (MSS).
Palantir's Maven Smart System is orders of magnitude more capable than the original computer-vision platform. MSS integrates feeds from multiple intelligence disciplines — imagery intelligence (IMINT), signals intelligence (SIGINT), human intelligence (HUMINT) reports, open-source intelligence (OSINT), and moving-target indicator (MTI) radar — into a single operating picture. It then uses a combination of trained AI models to correlate these signals, detect patterns consistent with military targets, and surface strike recommendations to human operators with associated confidence scores and target packages.
The Kill Chain, Compressed
Traditional targeting in the U.S. military runs through what doctrine calls the F2T2EA cycle: Find, Fix, Track, Target, Engage, Assess. In a deliberate targeting process — planning strikes against fixed installations over days — this cycle can take 72 hours or more, involving teams of intelligence analysts, judge advocates, commanders, and sometimes National Command Authority approval. In time-sensitive targeting against fleeting targets, the cycle has historically run 30 to 90 minutes.
With Maven Smart System, that cycle is being crushed. Pentagon briefers have described operators using MSS with a sequence that sounds almost mundanely simple: "Left click, right click, left click." The first click selects a proposed target surfaced by the AI. The second displays the target package, confidence scoring, and supporting intelligence. The third initiates the strike authorization request. Humans remain in the loop — technically — but the cognitive load, and therefore the deliberation time, has been radically reduced.
"We've gone from identifying the target, to now coming up with a course of action, in seconds. That's the capability Maven gives us."
— U.S. Military Official, briefing reporters at a Palantir demonstration eventIn Iran, this compressed kill chain has reportedly enabled the volume of strikes that would otherwise be operationally impossible. Eleven thousand targets across a country of 85 million people, struck over several months, represents a targeting tempo that no historical human-only process could sustain. Maven is why that tempo is possible.
Palantir's Role: Embedded Contractor on the Kill Chain
What makes the Palantir arrangement unique — and controversial — is the depth of integration. This is not a traditional defense contractor delivering a finished system and walking away. Palantir operates under what industry insiders call an "embedded contractor" model. Palantir engineers, referred to as "forward-deployed engineers," work alongside military units, often inside operational command structures, iterating on the software in near-real-time based on battlefield feedback.
This model was developed through years of exercises with the U.S. military under a program called Operation Scarlet Dragon — a series of live-fire exercises spanning multiple states and service branches in which Palantir engineers were physically co-located with warfighters. By the time the Iran conflict began, Palantir and the military had essentially merged their working processes for the targeting function.
Scarlet Dragon began as a small exercise using live ammunition and grew into a full multi-state, multi-branch program with Palantir forward-deployed engineers embedded alongside operators. It served as the proving ground for the Maven Smart System's warfighting integration model — a model now being tested at scale in the Iran theater.
Palantir's CEO Alex Karp has been characteristically unambiguous about what the company is doing and why. At a 2026 investor event, Karp argued that the West's ability to compress the kill chain with AI represented a civilizational-level capability advantage. "This is a have, have-not world," he said, arguing that a system compressing the kill chain from hours to seconds made an adversary "obsolete." He has framed Palantir's defense work not as arms dealing but as the defense of Western democratic values against authoritarian adversaries.
It's a frame that sits uncomfortably alongside the reported civilian toll in Iran, but it reflects the genuine belief inside Palantir and, increasingly, inside the Pentagon itself.
The Lumberjack Test: Maven Gets Its Autonomous Drone Upgrade
In the first week of April 2026, the Army disclosed details of a significant test conducted during Operation Lethal Eagle: Northrop Grumman's new Lumberjack drone conducting autonomous target detection using the Maven Smart System. This is a significant escalation in Maven's capability profile.
Lumberjack is a long-range, high-speed attritable drone — designed to be cheap enough to lose, capable enough to matter. It can carry both ISR payloads and munitions. In the Lethal Eagle test, Lumberjack was equipped with Maven's computer vision and target recognition algorithms, allowing it to autonomously detect, classify, and track targets without a human in the sensor loop at that stage of the process. The human authorization step occurs downstream — but the target identification is now fully machine-driven.
This is the progression that weapons researchers have been warning about for years. Each increment of autonomy is presented as a modest, reasonable step. A drone that autonomously detects targets is not a fully autonomous lethal system — a human still pulls the trigger. But the detection-to-strike timeline gets shorter with each generation, and the cognitive burden on the human authorizer gets lighter. At some point in this progression, the question of whether there is "meaningful human control" becomes genuinely difficult to answer.
Every increment of AI autonomy in the kill chain is justified individually as reasonable. Cumulatively, they shift the locus of effective decision-making from humans to machines. The Lumberjack/Maven integration is the latest step in a trajectory that has no clear stopping point defined in law, doctrine, or policy.
The Accuracy Crisis: When Maven Gets It Wrong
The most damaging stories emerging from the Iran war are not about Maven's capabilities — they are about its failures. In late March 2026, The Guardian published an investigation into a strike on what AI systems had identified as a military logistics hub. It was a school. The aftermath, including the deaths of civilians and a subsequent Pentagon investigation, threw Maven's accuracy claims into sharp relief.
The Pentagon and Palantir both declined to comment on Maven's overall performance statistics in Iran, citing operational security concerns. But independent analysts and former intelligence officers have raised a consistent set of concerns:
Pattern-of-Life Misclassification
Maven's AI models were trained predominantly on historical data from theaters like Iraq, Afghanistan, and Syria. The adversarial environment in Iran is structurally different: denser urban areas, more sophisticated Iranian operational security practices, civilian infrastructure that deliberately mimics military signatures, and electronic warfare capabilities designed to confuse AI sensor systems. Models trained on one type of conflict do not automatically generalize to another.
Confidence Score Gaming
Several intelligence officials, speaking anonymously to Bloomberg and the Washington Post, have raised concerns about how confidence scores generated by Maven are being used in practice. The system surfaces a target with a confidence score — say, 87% — that a given location is a military asset. Under pressure of operational tempo, operators are approving strikes at lower confidence thresholds than the system's designers intended. The threshold that seemed appropriate in a deliberate targeting exercise looks different when you've been awake for 36 hours and the operational tempo is 200 strikes a day.
The Speed-Accuracy Trade-Off
This is perhaps the most structurally intractable problem. The entire value proposition of Maven is speed. But accuracy in targeting fundamentally requires time — time to corroborate intelligence, to check multiple sources, to allow for competing analysis. When the system's primary virtue is compressing that deliberation time, you are by definition trading accuracy for speed. In a theater like Iran, where the population is dense and the adversary is sophisticated, that trade-off has consequences measured in civilian lives.
The Ethics Debate: Silicon Valley's Split on War
The use of Project Maven has reopened the fractures inside the American technology industry that first appeared in 2018 when Google employees protested the original Maven contract. Eight years later, the debate is louder, more urgent, and more consequential — because the system is no longer experimental. It is operational. It is at war.
Anthropic, the AI safety company founded by former OpenAI researchers, was reportedly excluded from government AI contracts in early 2026 following a dispute over what restrictions the company placed on how its models could be used. Anthropic had maintained usage policies prohibiting its AI from being used for fully autonomous weapons and certain domestic surveillance applications. The Trump administration's Pentagon viewed those restrictions as disqualifying. Anthropic was out.
The incident crystallized a growing fault line in the defense-AI ecosystem. On one side: companies like Palantir, Anduril, and Shield AI that have built their businesses explicitly around military applications and accept the operational requirements of warfighting without civilian-use restrictions. On the other: companies like Anthropic, and to a lesser extent OpenAI, that have attempted to maintain some version of an acceptable-use framework that limits the most autonomous and lethal applications.
The market is delivering a verdict. As of April 2026, Anduril's valuation has soared past $30 billion. Palantir's government revenue is at record highs. The companies willing to go to war without restrictions are winning the defense AI market. The companies with ethical guardrails are being pushed out of it.
"Frontier AI systems are simply not reliable enough to power fully autonomous weapons — but the military is deploying them anyway, with human authorization steps that are increasingly ceremonial rather than meaningful."
— AI Safety Researcher, Malwarebytes Security Podcast, April 2026International Law and the Targeting Responsibility Gap
Under international humanitarian law (IHL), the laws of war require that strike decisions meet three core criteria: distinction (distinguishing combatants from civilians), proportionality (the expected civilian harm must not be excessive relative to military advantage), and precaution (all feasible steps must be taken to verify the target and minimize collateral damage).
Project Maven's integration into the kill chain creates a genuine legal accountability gap. When a human operator approves a target recommended by an AI system — using a compressed deliberation process, with pre-generated target packages and pre-assessed collateral damage estimates produced by that same AI — is the human actually exercising the independent judgment required by IHL? Or are they rubber-stamping a machine decision?
Legally, the answer is probably "yes, it's lawful" — because there is a human making the formal authorization decision. But the spirit of meaningful human control, which IHL implicitly requires, is being hollowed out. The human is in the loop. Whether the human is controlling the loop is a different question.
The International Institute for Strategic Studies (IISS) published a major assessment this week on the proliferation of AI-enabled military technology in the Middle East, warning that existing governance frameworks are failing to manage the use of commercial AI providers in lethal applications. The IISS called explicitly for strengthened international norms requiring "meaningful human control over lethal decisions" — a phrase that has become the contested center of gravity in every legal and ethical debate about autonomous weapons.
The Agentic Future: What Comes After Maven
Even as Maven Smart System operates at combat scale in Iran, the Pentagon is already planning its successor. The buzzword at the DoD in early 2026 is "agentic AI" — AI systems that don't just analyze and recommend, but autonomously plan, coordinate multi-step operations, and execute tasks across complex environments without step-by-step human direction.
A veteran-founded startup, profiled by Defense One this week, debuted what it calls an agentic AI assistant specifically designed for warfighting — a system that can coordinate across multiple operational domains (air, ground, maritime, cyber, space) and generate integrated operational plans autonomously. The company is positioning itself explicitly for the wave of defense contracts it expects as the Pentagon moves from Maven-style recommendation systems to full agentic planning platforms.
This represents a qualitative leap. Maven recommends targets and compresses the targeting cycle. An agentic warfighting AI would potentially plan entire campaigns — selecting objectives, allocating assets, timing strikes, adapting to adversary actions — with human commanders reviewing and approving at a strategic level rather than a tactical one. The speed advantage would be enormous. The accountability questions would be even larger.
Project Maven Founded
Pentagon launches Algorithmic Warfare Cross-Functional Team. Google wins initial contract for computer vision applied to drone video feeds.
Google Withdraws Under Employee Pressure
Mass employee protests force Google's exit from Maven. Microsoft, Amazon, and Palantir step in. Palantir ultimately wins the primary development contract.
Maven Smart System Deployed in Training Exercises
Operation Scarlet Dragon and follow-on exercises integrate MSS with forward-deployed Palantir engineers embedded in military units.
Iran War: Maven Goes to War at Scale
MSS used as primary targeting support platform. 11,000+ targets struck. Accuracy questions emerge after civilian casualties. Anthropic excluded from DoD contracts over usage restrictions.
Lumberjack/Maven Autonomous Detection Demonstrated
Army discloses Operation Lethal Eagle test: Northrop Grumman Lumberjack drone conducting autonomous target detection via Maven Smart System. Next phase of autonomy integration underway.
What Should We Make of All This?
Project Maven is not a hypothetical. It is not a think-tank scenario or a policy paper. It is a live system that has participated in the targeting of more than 11,000 locations across Iran in 2026. The kill chain has been compressed from hours to seconds. Drones are now autonomously detecting targets using the same AI that recommends strikes to human operators. The companies with ethical guardrails have been pushed out of the market. The companies without them are thriving.
This is where we are. And the trajectory is not toward more human control — it is toward less. Every new capability demonstrated by Maven or its successors will be deployed, because the operational advantage is too large to ignore in a conflict environment. The Lumberjack/Maven integration demonstrated this month will become standard. The agentic planning systems being developed by defense AI startups will follow. Each step will be justified individually as reasonable, lawful, and controlled. Collectively, they will constitute a transformation of warfare that no international treaty, law, or norm is currently equipped to govern.
The question is not whether this is happening. It is happening. The question is whether the institutions designed to constrain the conduct of war — international humanitarian law, congressional oversight, military doctrine — can evolve fast enough to keep the accountability gap from becoming an abyss.
Based on what this week's reporting reveals, the answer is: not yet.
Democracy Now / Bloomberg (Project Maven Palantir Targeting Report) · The Guardian (Iran School Strike Investigation) · DefenseScoop (Army Lumberjack/Maven Lethal Eagle Test) · Defense One (Agentic AI Warfighting Startup) · IISS (AI-Enabled Military Technology Middle East Assessment) · Stanford HAI (Who Decides How America Uses AI in War) · Malwarebytes Security Podcast (Killer Robots and Autonomous Weapons)