America's First AI War:
How Autonomous Systems
Are Fighting in Iran

Section 1 // The Confirmation The Threshold Has Been Crossed

(U) On March 26, 2026, NPR published a report that will define how historians write about this era of warfare: "America's first AI-fueled war is unfolding right now in Iran." The statement was not speculative. It was not prefaced with caveats about the future of military technology. It was a present-tense declaration, sourced from ongoing military operations, confirmed by multiple senior personnel with direct knowledge of the conflict.

(U) This assessment treats that confirmation as the foundational datum from which all downstream analysis flows. We assess with HIGH CONFIDENCE that artificial intelligence systems are actively participating in targeting decisions, battle damage assessment (BDA), operational logistics optimization, and at least some degree of autonomous engagement sequencing in the Iran theater of operations. The age of AI-assisted warfare is not approaching. It has arrived.

(U) Reporting indicates that Anthropic's large language model Claude was initially employed in targeting-adjacent operational planning functions, a detail confirmed by The Hill. This single data point detonated an ethical and commercial shockwave throughout Silicon Valley and the broader AI industry that continues to reverberate at time of publication. Anthropic, upon learning the specific nature of its system's deployment — or prospective deployment — in lethal targeting chains, refused to authorize continued use for autonomous weapons applications. The Pentagon did not debate the point. It blacklisted Anthropic and immediately turned to OpenAI, which stepped into the role with apparent readiness.

(U) The speed of that substitution is itself a strategic signal. The fungibility of large language models in military applications — the ease with which one provider can be replaced by another — has collapsed whatever leverage AI companies believed they held over the ethics of their own technology's deployment. The market for compliant AI, it turns out, is deep and competitive. The market for principled refusal is a market of one, and the Pentagon has no obligation to shop there.

Section 2 // Order of Battle The AI Systems in the Iran Theater

(U) Reconstructing the AI order of battle from open-source reporting requires integrating confirmed disclosures, logical inference from documented procurement, and triangulation across independent sources. What follows represents this desk's current assessed picture of the AI architecture operating in or directly supporting Iran combat operations. Source reliability grades are assigned per standard intelligence practice.

// Palantir and the Maven Smart System

(U) Palantir Technologies holds the contract that matters most. Following Google's 2018 employee revolt over Project Maven — which this assessment addresses at length in Section 5 — the Defense Department pivoted its AI-ISR integration work to Palantir and Scale AI. The Maven Smart System (MSS), as it is now designated, is not the 2018 Google experiment. It is an eight-year-matured weapons-grade intelligence fusion platform that has absorbed the lessons of Ukraine, the refinements of adversarial machine learning research, and the operational feedback of near-peer competition.

(U) Multiple sources corroborate that the Maven Smart System is providing real-time targeting intelligence support in the Iran theater. This includes pattern-of-life analysis — the AI-assisted methodology by which a target's behavioral signature is mapped across time to establish kill-chain eligibility — as well as multi-INT fusion, which integrates signals intelligence, imagery, and human reporting into coherent, actionable targeting packages.

// The Replicator Program: Autonomous Strike at Scale

(U) The Replicator program — announced in August 2023 by Deputy Secretary of Defense Kathleen Hicks as a two-year initiative to deploy thousands of attritable autonomous systems — reaches its intended operational window precisely at the moment of the Iran conflict. We assess with MODERATE CONFIDENCE that Replicator-designated autonomous drone systems are either deployed in theater or available for immediate deployment. The program's explicit purpose — to counter China's military mass with autonomous numbers — has obvious dual-use applicability in any conventional conflict.

(U) Reporting has not, at time of publication, explicitly confirmed Replicator drone combat engagement in Iran. However, the convergence of timing, capability, and strategic incentive creates a strong inferential basis. The absence of confirmation may reflect classification, operational security, or the deliberate obfuscation of autonomous strike activities that could trigger legal and political scrutiny domestically and internationally.

Section 3 // Kill Chain Analysis How AI Has Compressed the Targeting Cycle

(U) The most consequential operational impact of AI integration in Iran is not the systems deployed — it is the time they have eliminated. Understanding this requires a brief orientation to the military targeting framework that AI has accelerated beyond prior recognition.

// The F2T2EA Framework

(U) The U.S. military's standard kinetic targeting cycle is structured around the F2T2EA framework: Find, Fix, Track, Target, Engage, Assess. In conventional pre-AI warfare, each phase demanded human analyst time, processing latency, and sequential handoffs between intelligence and operational elements. The cycle was measured in hours, sometimes days, for mobile or time-sensitive targets. A high-value target identified at dawn might not be engaged until the following morning, by which time it had moved, dispersed, or vanished into the population.

(U) AI has not merely accelerated individual steps in this cycle. It has collapsed the gaps between them. In the Iran theater, multiple sources indicate that the AI-assisted targeting cycle — from initial detection to targeting package delivery — has been compressed from the traditional 24 to 48-hour timeline to approximately 15 to 45 minutes for pre-characterized target types. For time-sensitive targets that fit established pattern-of-life profiles, the effective compression is even more dramatic.

(U) The practical consequences of this compression are not merely tactical. They are ethical. When a human targeting committee had 24 hours to deliberate a strike, there was time — however imperfectly used — for legal review, collateral damage estimation, political coordination, and cross-checking. When the same cycle runs in 20 minutes, that deliberative space evaporates. The question of who is actually authorizing lethal force — a human operator, a human reviewing an AI-generated recommendation, or an AI system operating within pre-authorized engagement parameters — becomes genuinely ambiguous.

// The Automation Gradient

(U) The Department of Defense officially maintains a policy requiring "appropriate levels of human judgment" in lethal targeting decisions. DoD Directive 3000.09, last updated in 2023, mandates that autonomous weapon systems must be designed to allow commanders to exercise "appropriate levels of human judgment over the use of force." The directive carefully avoids requiring human judgment at every targeting step, leaving substantial operational latitude.

(U) Reporting indicates that in practice, the AI systems operating in Iran function on what this desk characterizes as an automation gradient — a spectrum from human-in-the-loop, where a human explicitly approves each engagement, through human-on-the-loop, where a human can override but the system executes autonomously by default, toward human-out-of-the-loop, where pre-authorized engagement parameters allow the system to engage without real-time human involvement. Multiple sources suggest Iran operations are operating at the human-on-the-loop level for certain target categories, with possible human-out-of-the-loop authority for specific autonomous systems operating in contested airspace.

Section 4 // The Ethical Collision Anthropic, OpenAI, and the Fungibility of Conscience

(U) The collision between the Pentagon and Anthropic represents the first publicly documented instance of a major AI laboratory refusing military targeting use of its systems — and being immediately replaced by a competitor. The sequence of events, as reconstructed from available reporting, is as follows.

(U) Reporting indicates that Anthropic's Claude was employed, or considered for deployment, in targeting-adjacent operational planning functions within the Iran conflict architecture. The Hill confirmed Anthropic's involvement and subsequent refusal. Anthropic's position, consistent with its published acceptable use policies, prohibits use of Claude for autonomous weapons systems and lethal targeting chains. When the Pentagon's operational requirements crossed that line, Anthropic declined to continue. The Pentagon blacklisted the company. OpenAI, which has been progressively relaxing its military-use restrictions since late 2024, filled the role.

(U) The fungibility dynamic deserves analytical emphasis. The AI industry has largely operated under an implicit theory that ethical constraints imposed by individual companies carry operational weight — that a refusal by a major AI developer constitutes a meaningful check on military AI deployment. The Iran episode has empirically tested that theory and found it wanting. As long as multiple capable AI providers exist, and as long as government procurement can substitute between them within operational timelines, any single company's ethical stance becomes a market friction rather than a meaningful constraint.

(U) This creates a competitive race-to-the-bottom dynamic. The company that maintains the strictest ethical constraints loses government contracts to the company with fewer constraints. The company with fewer constraints gains market share, revenue, and the political capital that comes with being a defense partner. Over time, the structural incentive is for ethical constraints to erode across the industry, not to generalize from principled holdouts.

// The Engineers' Response

(U) The human dimension of this ethical collision has not been passive. Multiple sources report that engineers and researchers at both OpenAI and Google DeepMind have filed amicus briefs and internal objections regarding their companies' expansion into military AI applications. The echoes of the 2018 Google Maven revolt — which ultimately forced Google's withdrawal from the original Project Maven contract — are audible. Whether that historical precedent will repeat is assessed as unlikely given the changed structural environment.

(U) The 2018 environment featured a Google that was commercially dominant, less reliant on government contracts, and operating in a pre-competitive AI landscape where the cost of principled refusal was manageable. The 2026 environment features multiple AI companies competing intensely for government and defense contracts worth billions of dollars annually. The financial stakes of principled refusal have increased by an order of magnitude. Individual engineers filing objections operate within institutions whose financial incentives point in the opposite direction.

(U) The broader media environment has registered the ethical collision clearly. Reuters, The Guardian, CNBC, Al Jazeera, and The Hill are all actively covering the intersection of AI industry ethics and Iran combat operations. The public narrative is forming. Whether that narrative translates into regulatory or legislative pressure is a political question beyond the scope of this assessment, though we note that the pace of AI deployment in military operations has historically outrun the pace of governance response by years.

Section 5 // Historical Context Four Step Functions in the History of Killing

(U) Military history is not a smooth progression. It is punctuated by discontinuities — moments when a new technology or doctrine produces a step-function change in how wars are fought and won. The current moment is best understood in the context of its predecessors. We assess four major discontinuities in U.S. military operations over the past four decades, each of which set conditions for the current AI-warfare threshold.

(U) Each of these discontinuities generated a similar social and political response: initial alarm, subsequent normalization, and eventual institutionalization into doctrine and procurement. The precision revolution was once controversial; it is now assumed. Drone warfare was once a subject of intense ethical debate; it is now standard practice across dozens of state and non-state actors. AI warfare will follow the same arc. The debate about whether AI should be in the kill chain is already being overtaken by the fact that it is.

Section 6 // Project Maven Eight Years From Revolt to Combat

(U) Project Maven's trajectory from 2018 Google controversy to 2026 Iran combat deployment is one of the defining institutional stories of the AI age. This desk's analysis draws substantially on the reporting underpinning Katrina Manson's Bloomberg book, coverage of which aired on NPR on March 25 and 26, 2026 — the same reporting cycle that confirmed AI's role in Iran operations.

(U) Project Maven was initiated in 2017 under Deputy Secretary of Defense Robert Work as the Algorithmic Warfare Cross-Functional Team. Its stated mission: apply computer vision and machine learning to the DoD's massive and growing backlog of drone surveillance footage. The intelligence community was drowning in video data it lacked the human analysts to process. AI promised to automate the object detection, behavioral tagging, and pattern-of-life analysis that had previously required teams of human imagery analysts working around the clock.

(U) Google was the initial contractor. In 2018, that relationship became the most visible internal conflict in Silicon Valley's relationship with the defense sector. Over 3,000 Google employees signed a petition protesting the company's participation in what they characterized as the development of AI weapons technology. Several senior engineers resigned. The protest was effective: Google declined to renew the Maven contract when it expired in 2019.

// The Pivot to Palantir and Scale AI

(U) What the 2018 revolt did not accomplish was stopping Project Maven. The program pivoted to contractors without Google's employee sensitivity to military applications: primarily Palantir Technologies, Scale AI, and a constellation of defense-specific AI companies. Without the cultural friction of a consumer technology company workforce, these contractors built faster, integrated more deeply with operational systems, and accepted requirements that Google's workforce would have found intolerable.

(U) The eight years between 2018 and 2026 were not wasted. Reporting indicates the Maven system was tested in classified exercises, refined against imagery datasets from Syria, Afghanistan, and the Ukraine theater, and progressively integrated into the Joint All-Domain Command and Control architecture. The Google revolt delayed Maven by approximately 18 months in our assessment. It did not change its destination.

// Scale AI and the Training Data Infrastructure

(U) Scale AI's role in the Maven ecosystem deserves separate treatment. The company's core business — providing human-labeled training data for machine learning systems — is the foundational infrastructure layer beneath all AI targeting capability. An AI system that can reliably distinguish a mobile missile launcher from a fuel truck, or identify human behavioral patterns consistent with weapons emplacement, is only as capable as the labeled training data used to train it. Scale AI has spent years building the datasets and annotation infrastructure that make military AI viable at operational scale.

(U) The integration of Scale AI's data infrastructure with Palantir's operational platform and the DoD's sensor networks creates a closed-loop system capable of continuous improvement: new operational data from Iran can, in principle, be fed back into retraining pipelines, making the targeting AI progressively more accurate against Iranian-specific target signatures, terrain, and tactics. This self-improving capability is among the most strategically significant and least publicly discussed aspects of the deployed system.

Section 7 // JADC2 and Systems Integration The Command Architecture of AI Warfare

(U) The AI systems described in preceding sections do not operate as isolated tools. They are nodes in an integrated command architecture: the Joint All-Domain Command and Control system, known as JADC2. Understanding JADC2 is essential to understanding how AI warfare functions at the operational level, as opposed to the individual system level.

(U) JADC2 is the DoD's conceptual and technical framework for connecting sensors, shooters, and decision-makers across all domains — land, sea, air, space, and cyber — in a unified, AI-assisted information environment. Its ambition is to allow any sensor in the joint force to feed any shooter, with AI systems managing the information flow, prioritizing targeting opportunities, and presenting commanders with decision-ready packages rather than raw data.

(U) In the Iran theater, JADC2-integrated operations mean that a signal intercept from a National Security Agency sensor, a satellite image from the National Geospatial-Intelligence Agency, a pattern-of-life track from a Reaper drone, and a human report from a ground asset can be fused by the Maven Smart System into a single targeting package within minutes. A commander reviewing that package is not weighing fragmentary intelligence from disparate sources. They are reviewing an AI-synthesized assessment that already integrates, weights, and presents the most operationally relevant conclusion.

(U) The cognitive implications of this architecture are subtle but profound. When a commander approves a target package, they are not making an independent analytical judgment. They are approving an AI judgment that has been presented in a format optimized for human acceptance. The human is in the loop, technically. But the loop's direction of influence has inverted: the AI is leading the human toward a decision rather than supporting one the human has independently reached.

Section 8 // Cyber AI Operations The Invisible Front

(U) Open-source confirmation of AI-assisted cyber operations against Iran is, by the nature of the domain, substantially more limited than confirmation of kinetic AI operations. This desk nonetheless assesses with MODERATE-LOW CONFIDENCE that AI is playing a supporting role in U.S. Cyber Command and NSA offensive cyber operations against Iranian infrastructure, military networks, and industrial control systems.

(U) The precedent for U.S. offensive cyber operations against Iran is extensive and publicly documented. The Stuxnet operation against Iranian nuclear centrifuges, attributed to U.S. and Israeli intelligence services, demonstrated more than a decade ago that cyber operations could produce physical effects on Iranian infrastructure. Subsequent operations, less fully documented in open sources, suggest a persistent U.S. cyber presence in Iranian networks.

(U) AI integration in offensive cyber operations enables capabilities that were previously impractical: automated vulnerability discovery at machine speed, adaptive malware that can modify its behavior based on network responses, AI-assisted social engineering for credential harvesting, and autonomous lateral movement through compromised networks. The application of these capabilities against Iranian military command networks, air defense systems, and critical infrastructure represents a warfare domain that operates largely below the threshold of public visibility.

Section 9 // Strategic Implications What Iran 2026 Means for the Future of War

(U) This assessment's final analytical section addresses the strategic implications of America's first AI war for the future of armed conflict globally. We identify four primary implications that this desk assesses as high-confidence directional trends, acknowledging that specific timelines and magnitudes carry greater uncertainty.

// Implication One: The Proliferation Imperative

(U) Every major military power — China, Russia, Israel, the United Kingdom, France, and India — is watching the Iran conflict with the intensity of a live technological demonstration. Each will draw conclusions about the operational effectiveness of AI-integrated targeting. Each will accelerate its own programs accordingly. The Iran conflict will do for military AI what Desert Storm did for precision weapons: validate the concept in combat and trigger a global proliferation race. Adversary AI targeting programs that are currently in development or testing will reach operational status faster as a result of the Iran data. The window in which the United States holds a meaningful AI warfare advantage is measured in years, not decades.

// Implication Two: The Governance Gap Widens

(U) International humanitarian law — the laws of armed conflict — was developed in an era when the relevant actors were humans making decisions with human cognitive processes. The Geneva Conventions, the Additional Protocols, the targeting principles of distinction, proportionality, and precaution all assume a human decision-maker capable of exercising moral judgment. AI targeting systems compress decision timelines below the threshold at which meaningful legal review can occur, distribute responsibility across systems and operators in ways that obscure accountability, and make targeting recommendations that no individual human fully understands. The legal architecture governing war is being outpaced by the technology of war. Efforts to negotiate international AI weapons governance treaties are at least a decade behind the operational reality on the ground in Iran today.

// Implication Three: The Commercial-Military Entanglement Deepens

(U) The Anthropic-OpenAI substitution dynamic described in Section 4 has a structural implication beyond the individual episode: the United States military is now dependent on commercial AI infrastructure for core warfighting functions in a way that has no historical precedent. The Department of Defense does not own the underlying models. It does not fully control the training pipelines. It is operationally dependent on companies whose primary governance is by private boards, venture capital investors, and competitive market dynamics. This creates a novel vulnerability: the AI systems fighting America's wars can be withdrawn, updated, or compromised by corporate decisions made in San Francisco.

// Implication Four: The Threshold Has No Return

(U) Once AI targeting is deployed in combat at scale and produces operationally superior results, the political and military cost of removing it becomes prohibitive. Commanders who have experienced 20-minute targeting cycles will not voluntarily return to 24-hour cycles. Institutions that have optimized their doctrine, training, and organizational structure around AI-assisted operations will resist reverting to pre-AI architectures. The Iran conflict, if it produces the targeting outcomes the deployed systems are designed to achieve, will lock in AI warfare as the baseline for all future U.S. military operations. This is not a prediction. It is an observation about institutional momentum and the irreversibility of operational precedents.