The Integration of Artificial Intelligence in Modern Kill Chains: Systems, Doctrines, and Strategic Implications

Full report

On this page

The character of modern warfare is undergoing a fundamental and irreversible transformation, driven primarily by the integration of artificial intelligence (AI) and machine learning (ML) into the military targeting process. At the core of this transformation is the "kill chain," a doctrinal military concept that defines the structural sequence of an attack1. Historically, this integrated, end-to-end loop operated at "human speed," requiring hours, days, or even weeks for intelligence analysts and commanders to identify threats, allocate resources, and authorize strikes3. The advent of the AI kill chain compresses this entire sequence into mere seconds3. By leveraging computer vision models to scan satellite imagery, sensor fusion algorithms to combine radar and signals intelligence, and natural language processing to sift through intercepted communications, AI systems can rapidly score potential targets and present actionable data to human operators or directly to autonomous weapons platforms3. The strategic advantage in modern conflict increasingly belongs not necessarily to the force with the most exquisite kinetic weapons, but to the force that can execute the kill chain at the fastest speed, widest scale, and highest degree of survivability5. This report provides an exhaustive technical and strategic analysis of the weapons systems, command and control architectures, and algorithmic frameworks that utilize AI kill chains. It examines the mechanisms enabling machine-speed warfare, from advanced Radar Automatic Target Recognition to distributed swarm consensus algorithms, and explores the doctrinal shifts and profound legal challenges inherent in automating the application of lethal force.

Doctrinal Frameworks: F2T2EA, D3A, and the UAS Kill Chain

To understand how AI is integrated into modern weapons systems, it is essential to analyze the doctrinal frameworks that govern targeting. The United States military and allied forces primarily utilize three overlapping models to conceptualize and execute strikes: F2T2EA, D3A, and the OODA loop, all of which are being heavily modified by artificial intelligence. The F2T2EA kill chain is a dynamic targeting process consisting of six distinct phases, described as a "chain" because an interruption at any stage can disrupt the entire process1. In the Find phase, intelligence, surveillance, and reconnaissance assets detect a potential target1. AI enhances this phase through automated object detection across wide surveillance areas, frequently utilizing computer vision to identify anomalies in satellite feeds2. The Fix phase involves identifying and classifying the detected contact to obtain specific coordinates with sufficient fidelity for engagement2. Here, AI classification models and multi-modal sensor fusion algorithms excel at confirming target identity against background clutter1. The Track phase requires monitoring the target's movement continuously2. Machine learning predictive kinematics assist in maintaining tracks even when sensor contact is temporarily lost or degraded by adversary electronic warfare1. During the Target phase, command and control capabilities assess the value of the target, select the appropriate weapon system to achieve the desired effect, and evaluate legal constraints such as rules of engagement1. AI optimization algorithms are increasingly deployed to solve these complex weapon-target pairing equations instantaneously4. The Engage phase involves the physical application of the weapon to the target, a stage where AI terminal guidance takes over in autonomous munitions to ensure precise impact despite communication degradation2. Finally, the Assess phase evaluates the effects of the attack, where AI image clustering models analyze post-strike data to confirm target neutralization or recommend re-engagement2. While joint targeting relies heavily on the F2T2EA model, the United States Army maintains D3A—Decide, Detect, Deliver, Assess—as the doctrinal cornerstone of fires and effects integration at the brigade and division levels4. In the Decide phase, AI tools such as game theory models, decision trees, and logistic regression support enemy course-of-action development and attack asset prioritization4. In the Detect phase, AI executes pattern association and anomaly detection to functionally characterize targets at massive scales4. The Deliver phase utilizes prescriptive analytics to refine weapon-target pairings, while the Assess phase leverages explainable AI tools to assist in image interpretation and effects validation4. Both of these models fit within the broader Observe, Orient, Decide, Act (OODA) loop, a cyclical framework where AI profoundly impacts the Orient and Decide phases by fusing massive datasets to provide situational awareness that far exceeds human cognitive limits6. The proliferation of Uncrewed Aerial Systems (UAS) has introduced new complexities to these traditional models, necessitating specific UAS kill chains. A UAS comprises three main components: the sensor, the controller, and the communication links, each presenting unique vulnerabilities1. While traditional air defense kill chains target the aircraft, payload, or logistical backbone, the UAS kill chain introduces three additional target options: the control element, the human operator, and the control link mechanism1. Suppressing a UAS mission can often be achieved without physically destroying the aerial vehicle by targeting these specific vulnerabilities. For instance, sensors are vulnerable to spoofing of Global Positioning System, GLONASS, Galileo, and BeiDou signals, while communication links are vulnerable to video and telemetry downlink interception or jamming1. AI systems are increasingly deployed to automatically detect these specific emissions and execute non-kinetic electronic warfare engagements to break the adversary's UAS kill chain at the earliest possible stage1.

Algorithmic Foundations: RATR, WTA, and Swarm Consensus

The efficacy of an AI kill chain is dictated by the underlying mathematics, physics models, and algorithms that process raw data into lethal action. Three highly specialized technical domains enable machine-speed warfare: Radar Automatic Target Recognition, Weapon Target Assignment optimization, and Distributed Swarm Consensus algorithms.

Radar Automatic Target Recognition (RATR)

Optical sensors are frequently degraded by weather, darkness, and physical camouflage, forcing modern militaries to rely heavily on radar modalities. The algorithmic process of automatically detecting, classifying, and identifying targets within these returns is known as Radar Automatic Target Recognition (RATR)8. The information available for RATR depends heavily on the target's size parameter relative to the radar wavelength, categorized into three physics regimes. In the Rayleigh region, targets produce a point-scatterer response with limited structural discriminability8. In the Resonance (or Mie) region, frequency-dependent natural resonance poles allow for specialized identification techniques8. Finally, in the Optical region, physical-optics scattering-center theory applies, characterizing targets by sparse dominant scatterers, which is highly informative for targeting algorithms8. Modern RATR pipelines exploit diverse modalities, including Synthetic Aperture Radar (SAR), which delivers coherent two-dimensional reflectivity maps; Inverse SAR (ISAR), which images moving targets by exploiting cross-range Doppler resolution; High-Resolution Range Profiles (HRRP); and micro-Doppler signatures, which provide fine-grained time-frequency data to identify moving components like drone rotor blades8. Historically, SAR ATR relied on Constant False Alarm Rate detectors, which determine a threshold based on background noise to flag potential targets8. However, against complex clutter or stealthy targets, these statistical models generate high false alarm rates9. Consequently, contemporary RATR relies almost exclusively on deep learning architectures, specifically Convolutional Neural Networks, to process the complex-valued data (both magnitude and phase) of SAR imagery9. Because gathering massive, labeled SAR datasets of adversarial hardware is prohibitively expensive and restricted by privacy concerns—often relying on legacy datasets like the Moving and Stationary Target Acquisition and Recognition (MSTAR) collection from the 1990s—researchers have developed advanced techniques to train algorithms on limited data8. This includes self-supervised learning frameworks utilizing masked Siamese Vision Transformers, Generative Adversarial Networks for synthetic data generation, and hybrid quantum-classical architectures9. Recent studies indicate that in hybrid quantum models, magnitude-only encoding can achieve highly accurate SAR ATR classification without the heavy computational burden of phase-aware methods12. Despite these advancements, RATR neural networks remain vulnerable to adversarial attacks; the introduction of geometric perturbations or adversarial thermal camouflage patches can drastically reduce the algorithm's average precision, effectively breaking the Find and Fix phases of the kill chain12.

Weapon Target Assignment (WTA) Optimization

Once targets are identified and validated, the military commander faces the Weapon Target Assignment problem, a fundamental defense operations research challenge15. The objective of WTA is to optimally assign a specific number of weapons to a specific number of targets to minimize the expected post-engagement survival value of the targets, or conversely, to maximize the survival of friendly assets15. Because the variables scale exponentially with the number of assets, WTA is recognized mathematically as an NP-complete combinatorial optimization problem15. Traditional command and control systems approach WTA using centralized planners executing Mixed Integer Linear Programming or heuristic algorithms such as genetic algorithms, simulated annealing, and multi-objective particle swarm optimization15. In these traditional models, algorithms must account for strict constraints, including resource limitations, launch delays, and physical or seeker interference18. To prevent two interceptors from colliding or their radar seekers from blinding one another, traditional MILP formulations discretize the time window to generate Predicted Intercept Points, pre-calculating interference tables before the solver executes the engagement18. However, in dynamic, high-attrition environments, centralized planners create a single point of failure. If a weapon is destroyed mid-flight by adversary air defenses, a centralized system struggles to recalculate assignments in real-time, resulting in wasted munitions attacking already destroyed targets or ignoring high-priority threats17. To resolve this, modern AI frameworks are deploying distributed, autonomous assignment planning. By utilizing continuous convex relaxation of the associated cost functions and decentralized primal-dual optimization algorithms, individual weapons can compute their own assignments autonomously, updating their targeting priorities on the fly even under intermittent communication conditions17. Furthermore, cutting-edge research is integrating Large Language Models into the WTA framework to inject generalized artificial intelligence into cooperative missile guidance19. Rather than relying on rigid, pre-defined weighting parameters, the LLM treats tactical decision-making as a contextual reasoning task19. By evaluating spatial geometries, closing velocities, threat directions, and doctrinal asset priorities, the system can dynamically adapt to asymmetric mission configurations, cooperatively allocating redundant interceptors to single, high-value targets to guarantee neutralization19.

Distributed Consensus and Swarm Algorithms

As military strategy shifts toward employing thousands of autonomous agents simultaneously, centralized control architectures become mathematically and physically untenable due to bandwidth limitations and electromagnetic jamming22. The solution is distributed swarm intelligence, leveraging consensus algorithms inspired by blockchain technology and distributed computing to maintain a unified operational picture across the swarm23. Algorithms such as SwarmRaft, an adaptation of the traditional Raft consensus protocol, enable decentralized Uncrewed Aerial Vehicle swarms to agree on critical state updates, including location and heading, even when Global Navigation Satellite Systems are completely denied by electronic warfare24. In the event of a GPS signal loss or sensor malfunction, the swarm utilizes sensor fusion and peer-to-peer data sharing to algorithmically reconstruct the blinded drone's position based on its last known state and trajectory, ensuring the swarm maintains coherence and fault tolerance without a central leader24. Other advanced frameworks, such as DiRAC and DANCeRS, utilize zone-partitioned architectures with dynamically elected zone leaders and Gaussian Belief Propagation to manage task assignment and real-time, collision-free path planning across thousands of robotic agents22. In both industrial settings and military environments, these decentralized multi-agent swarms form a robust "digital immune system." They utilize mechanisms like the Decentralized Task Authorization and Validation token system and Rotational Leadership Role voting to validate threats in sub-millisecond timeframes, ensuring that the swarm reacts organically and instantaneously to hostile engagements while remaining highly resistant to single points of failure25. Projects such as the Defense Advanced Research Projects Agency's Autonomous Multi-Domain Adaptive Swarms-of-Swarms (AMASS) seek to establish this type of common algorithmic language to command highly autonomous networks dynamically across air, land, and sea domains29.

Strategic Command and Control: The AI Target Generation Factory

The most profound and immediate application of the AI kill chain currently exists not in individual robotic platforms, but in hyperscaled, AI-enabled decision-support systems. These software platforms function as the central nervous system of modern warfare, aggregating intelligence from thousands of disparate sources to generate actionable target lists for human commanders at an unprecedented scale.

Project Maven and the Maven Smart System

Project Maven, initiated by the United States Department of Defense in 2017 under the title Algorithmic Warfare Cross-Functional Team, represents a historic paradigm shift in military intelligence workflows30. Originally conceived by Deputy Secretary Robert O. Work to apply computer vision to full-motion video from drones, the program has evolved into the Maven Smart System, the foundational AI-enabled software platform for Combined Joint All-Domain Command and Control30. Following the withdrawal of early partner Google in 2018 due to employee protests, data integration company Palantir became the primary industry partner, with its DOD contract ceiling for Maven surpassing one billion dollars in 202530. Operating under the National Geospatial-Intelligence Agency and transitioning to the Chief Digital and Artificial Intelligence Office, the platform integrates data across all combatant commands30. The Maven Smart System essentially collapses the sensor-to-shooter timeline into a single, synchronized graphical user interface30. Prior to its implementation, military analysts spent the vast majority of their time gathering data and manually transferring detections across multiple disparate systems34. Today, MSS integrates up to 179 distinct data sources across land, sea, air, space, and cyber domains, including satellite imagery, communications intercepts, infrared sensors, and synthetic-aperture radar30. The targeting workflow within MSS perfectly illustrates the realization of the AI kill chain. Computer-vision models automatically analyze incoming imagery, visually highlighting potential targets—such as enemy tanks or missile launchers—with yellow-outlined boxes, while blue-outlined boxes designate friendly forces and no-strike zones30. Once a human operator validates the AI detection, the system utilizes recommendation algorithms and Large Language Models to propose optimal courses of action31. The system discerns the nearest available weapons, calculates flying times, checks weapons loading details, and ensures compliance with doctrine30. Through components like the Target Workbench, analysts can sequence targets by priority and route execution data directly to artillery networks or aircraft via the Joint Range Extension Applications Protocol30. This integration was notably demonstrated during the Scarlet Dragon Oasis exercise, where MSS interfaced through JREAP-A to an Air Operations Center, directing a B-52 bomber to drop live ordnance on a training target30. The operational efficiency gained by this system is staggering. Assessments of the Scarlet Dragon exercises revealed that by using MSS, a targeting cell of only twenty personnel matched the performance of the Operation Iraqi Freedom time-critical targeting cell, which originally required over two thousand staff members, reducing the time to pass targeting data from twelve hours to under one minute36. During recent conflicts, such as the United States' response to attacks in the Red Sea and Yemen, MSS facilitated a tenfold increase in targeting capacity, supporting the identification and prosecution of over one thousand targets within the first twenty-four hours of engagement30. Furthermore, MSS features an open, extensible architecture designed to integrate third-party AI models, creating a comprehensive kill web35. Recent integrations include Safran.AI for highly accurate object classification, Quantum Systems' MOSAIC UXS for autonomous drone mission planning and routing, and Hadean's dominAI, which generates and simulates multiple courses of action simultaneously by fusing live data and Allied Joint Publications doctrine35.

The Gospel and Lavender: AI in Urban Combat Operations

While the United States emphasizes multi-domain integration, the Israel Defense Forces have aggressively operationalized AI target generation systems specifically tailored for high-density urban combat, prominently highlighted during operations in the Gaza Strip. Two distinct systems—"The Gospel" (Habsora) and "Lavender"—illustrate the bifurcation of AI targeting into infrastructure generation and personnel generation37. The Gospel is an AI-based decision support system strictly focused on identifying physical structures37. It operates as a massive information aggregator, fusing satellite imagery, drone footage, cyber intelligence, and human intelligence to locate buildings and command centers utilized by hostile actors38. By pattern-matching known enemy infrastructure, The Gospel automatically generates recommendations for airstrikes on private residences and military installations38. Crucially, the system calculates potential civilian casualties in advance, attaching a collateral damage score to each target file38. The speed of this system is unprecedented; while human analysts historically produced roughly fifty targets per year in Gaza, The Gospel allows the IDF to generate over one hundred new targets every single day, outpacing the military's physical capacity to conduct strikes38. Conversely, Lavender is an AI-powered database utilized to assign risk scores to individuals, evaluating the likelihood that a person is a member of groups such as Hamas or Palestinian Islamic Jihad39. Lavender utilizes semi-supervised machine learning, drawing on visual information, social media connections, cellular data, and movement patterns to flag suspicious features39. In the early stages of the 2023 conflict, Lavender reportedly flagged up to 37,000 Palestinian men as potential targets38. Lavender does not independently generate authorized kill lists, but rather serves as a central query system; when an analyst evaluates an individual, they query Lavender to instantly compile all known intelligence files39. These systems are heavily augmented by a supplementary automated tracking program known as "Where's Daddy?", which utilizes mobile phone geolocation to track marked individuals in real time and alert commanders the moment the target enters their family residence, facilitating nighttime strikes37.

International Humanitarian Law and the Ethics of Automation

The deployment of high-speed AI target generation systems, particularly in dense urban environments, has sparked intense global debate regarding compliance with the Law of Armed Conflict, specifically the principles of distinction, precautions in attack, and proportionality39. The principle of distinction mandates that attacks must be strictly limited to lawful military objectives and combatants39. Theoretically, by organizing complex datasets and weeding out irrelevant information, systems like The Gospel and Lavender should enhance the reliability of target identification and decrease the likelihood of erroneous attacks against civilians39. Furthermore, the customary international law obligation regarding precautions in attack requires militaries to take feasible steps to verify a target39. In high-tempo operations, utilizing AI to rapidly cross-reference intelligence reduces the risk of relying on stale data, leading some legal scholars to argue that failing to utilize such advanced verification systems could itself constitute a violation of legal duties39. Similarly, while AI does not legally calculate proportionality, by providing granular data on target locations, it equips human planners with the precise information needed to select tactics that mitigate excessive collateral damage39. However, investigative reports and humanitarian organizations suggest severe operational realities that contradict these legal theories. The primary risk is "automation bias," a phenomenon where human operators, overwhelmed by the high tempo of machine-speed warfare and institutional pressure to produce targets, act as mere rubber stamps for the algorithm39. Reports indicate that IDF analysts often spent as little as twenty seconds reviewing a Lavender-generated target before authorizing a strike, relying entirely on the machine's stated 90% accuracy rate39. When applied at scale, a known 10% error rate on 37,000 targets translates to approximately 3,700 individuals potentially misidentified due to shared names, shared cell phones, or coincidental communication patterns resembling known operatives39. Furthermore, the practice of algorithmically generating vast numbers of targets, combined with policy shifts allowing for significant collateral damage, strains the principle of proportionality39. Reports suggest that in pursuit of junior operatives identified by Lavender, the military authorized up to fifteen or twenty civilian casualties per strike as permissible collateral damage, often utilizing unguided "dumb bombs" that destroy entire apartment buildings39. For senior officials, the authorized civilian casualty threshold reportedly exceeded one hundred39. This realization of the "Dahiya Doctrine"—the use of overwhelming force against infrastructure to establish deterrence—illustrates the profound tension between the mathematical efficiency of the AI kill chain and the ethical mandates of international law39. When human oversight is reduced to a twenty-second verification of a target's gender to confirm a machine's output, the legal safeguards designed to protect non-combatants are effectively nullified39.

Autonomous Munitions and the Tactical Edge

While C2 systems orchestrate the strategic battlefield, the Engage phase of the kill chain is increasingly dominated by autonomous munitions capable of executing strikes without a continuous human command link. This shift is primarily driven by the realities of modern electronic warfare and the need for attritable mass.

Terminal Autonomy in the Ukraine Conflict

In high-intensity conflicts, such as the war in Ukraine, the proliferation of electronic warfare jamming and GPS spoofing frequently severs the communication links between drones and their human operators45. A traditional remotely piloted UAS relies on a continuous control link and video downlink to complete the kill chain; when these are jammed, the weapon is rendered useless1. To overcome this, both Russian and Ukrainian forces have rapidly developed munitions featuring onboard edge computing and terminal AI autonomy45. These systems represent a qualitative departure from remotely piloted expendable drones45. The Armed Forces of Ukraine have deployed systems like the Saker Scout drone, which utilizes advanced computer vision to autonomously detect, identify, and strike targets47. By processing optical data onboard, it bypasses the need for GPS or a human operator during its terminal dive46. Similarly, Russian forces have heavily utilized the ZALA Lancet loitering munition44. Recent iterations of the Lancet reportedly feature autonomous target recognition algorithms that allow the drone to identify specific armor profiles and execute an autonomous "reconnaissance-strike contour," selecting and engaging targets entirely independent of external communication44.

The Replicator Initiative and Attritable Mass

Recognizing the shifting balance of power and the necessity of mass in future conflicts, the United States Department of Defense launched the Replicator Initiative in August 202349. Championed by Deputy Secretary of Defense Kathleen Hicks and overseen by the Defense Innovation Unit, Replicator aims to overcome the DoD's notoriously slow acquisition cycle—the so-called "production valley of death"—by fielding multiple thousands of "attritable autonomous systems" across all domains within eighteen to twenty-four months49. The strategic logic behind the first iteration of Replicator is to directly counter the People's Republic of China's military mass by deploying intelligent, low-cost swarms in the Indo-Pacific region49. A key component of the first tranche is the AeroVironment Switchblade 600 loitering munition52. However, executing Replicator has proven challenging; technical issues regarding software integration, command and control, and high costs persist52. While Ukrainian drones procured for similar purposes can cost as little as $300, systems like the Switchblade 600 are estimated to cost over $100,000 per unit, complicating the vision of cheap, attritable mass52. Following the focus on offensive swarms, Defense Secretary Lloyd Austin announced the Replicator 2 initiative in September 2024, shifting the Pentagon's focus to Counter-Small Uncrewed Aerial Systems for the defense of critical military installations54. The proliferation of cheap, asymmetric drone threats necessitates AI-driven defensive kill chains that can detect and defeat incoming swarms faster than human operators can react56. A prominent capability acquired under Replicator 2, managed via the newly established Joint Interagency Task Force 401, is the DroneHunter F70055. Developed by Fortem Technologies, the DroneHunter utilizes an onboard AI-integrated radar to autonomously detect, track, and pursue hostile drones in complex environments55. Rather than using kinetic explosives which risk collateral damage to domestic or allied infrastructure, the DroneHunter engages by firing a tethered net to capture the enemy drone, safely towing it away for forensic analysis55. This represents a highly specialized, non-lethal AI kill chain optimized for point defense.

Strategic Postures: The United States vs. China

The integration of AI into the kill chain is not occurring in a vacuum; it is the focal point of geopolitical competition between the United States and the People's Republic of China. Both nations view artificial intelligence as the decisive factor in future warfare, yet their doctrinal approaches highlight distinct strategic philosophies and risk tolerances.

The Chinese PLA and "Intelligentized Warfare"

The People's Liberation Army views the character of warfare as evolving through three distinct historical stages: Mechanization, Informatization, and Intelligentization61. The PLA officially declared the achievement of mechanization—the ability to deliver basic firepower in the physical space—in 2020, and is actively pursuing informatization, which focuses on networking and C4ISR capabilities62. Currently, under the direct mandate of Xi Jinping, the PLA is aggressively transitioning toward Intelligentized Warfare (智能化战争)44. Chinese military strategists argue that while previous eras of war were fought in the physical and information domains, future conflicts will be decided in the "cognitive space"62. This doctrine envisions a new type of warfare where AI works side-by-side with commanders, providing battlefield perception systems that suggest optimal target sets, thereby achieving "human self-transcendence"62. The PLA refers to this dynamic as "algorithmic warfare," anticipating an arms race based strictly on computing capacity and decision-making speed62. Over time, PLA theorists expect frontline combatants to be gradually phased out and replaced entirely with intelligent swarms of drones operating collectively7. By leveraging the national strategy of military-civil fusion, the PRC aggressively integrates commercial AI innovations into unmanned intelligent combat systems, aiming to challenge United States dominance in the Western Pacific65. Evidence suggests these integrations are already altering battlefield effectiveness; AI-enhanced first-person-view drones reportedly achieve strike accuracy rates of 70-80%, compared to 10-20% for traditional systems44. However, U.S. analysts warn that PLA theorists may be fundamentally overestimating the maturity of current AI7. By pushing rapidly toward autonomous targeting, the PLA risks deploying brittle systems susceptible to catastrophic failure outside their highly structured training environments, falling victim to the "normal accidents" theory of complex, tightly-coupled system failures44.

The United States and Responsible AI Policy

The United States approach, characterized by the CJADC2 framework and the Replicator Initiative, seeks to network every sensor to every shooter to achieve decision dominance3. Unlike the PRC, the U.S. Department of Defense places a heavy, public emphasis on the ethical frameworks governing autonomous weapons, prioritizing transparency and accountability. Department of Defense Directive 3000.09 (Autonomy in Weapon Systems), originally issued in 2012 and significantly updated in January 2023, serves as the cornerstone of U.S. military policy66. The directive dictates that all autonomous and semi-autonomous systems must be designed to allow commanders and operators to exercise "appropriate levels of human judgment over the use of force"66. The policy establishes distinct categories based on the role of the human operator: lethal autonomous weapon systems, or "human-out-of-the-loop" systems, can select and engage targets without further intervention once activated67. This is contrasted with "human-on-the-loop" systems, where operators monitor and can halt engagements, and "human-in-the-loop" semi-autonomous systems, which only engage targets explicitly selected by an operator67. Directive 3000.09 does not ban lethal autonomous weapon systems, nor does it require strict manual human control for every microsecond of an engagement; rather, it mandates rigorous hardware and software verification, realistic operational testing against adaptive countermeasures, and strict adherence to the Law of Armed Conflict66. The 2023 update mandates that the design and deployment of AI systems must remain consistent with the DoD AI Ethical Principles—ensuring systems are Responsible, Equitable, Traceable, Reliable, and Governable66. Furthermore, any lethal autonomous weapon system capable of selecting and engaging targets must undergo specialized high-level reviews by the Undersecretary of Defense for Policy, the Undersecretary of Defense for Research and Engineering, and the Vice Chairman of the Joint Chiefs of Staff before formal development and fielding68. While civil society groups and human rights organizations frequently criticize the directive for containing waivers to these senior reviews and utilizing ambiguous language—such as replacing the legally binding term "shall" with "will" in the 2023 update—the policy nevertheless establishes a high international standard70. In contrast to the secrecy surrounding the autonomous weapon policies of adversaries, the DoD's public guidelines explicitly address the vulnerabilities of AI, noting that failures can arise from software coding errors, enemy cyber attacks, jamming, spoofing, and unanticipated battlefield situations53.

Legal, Ethical, and Operational Vulnerabilities

As AI kill chains mature toward full autonomy, they introduce profound risks that threaten to undermine operational stability and humanitarian law. Current Large Language Models and computer vision algorithms present unique vulnerabilities due to their underlying mechanics. The most pressing issue is the "black-box" problem. Deep learning models operate via statistical prediction and lack true comprehension of doctrinal terminology or contextual nuance4. When an AI generates a high-confidence threat score, it cannot easily explain the semantic, human-understandable reasoning behind that conclusion4. Under international law, commanders are legally responsible for assessing military necessity and proportionality4. Utilizing a black-box system creates a severe legal liability, as human operators cannot verify the logical steps taken by the algorithm prior to authorizing a strike, forcing developers to prioritize rigid traceability over opaque accuracy4. Furthermore, AI models are highly susceptible to adversarial exploitation. A slight alteration to a physical object—known as an adversarial patch—can cause a highly advanced radar targeting system to misclassify a tank as a civilian vehicle14. Data poisoning during the training phase, or electronic spoofing in the field, can fundamentally break the logic of the kill chain, creating scenarios where autonomous swarms misidentify friendly forces24. Finally, the pursuit of machine-speed warfare risks triggering "flash escalations," where opposing autonomous systems react to one another's behaviors in an uncontrollable algorithmic feedback loop, deploying lethal force before human diplomats or commanders possess the time to intervene3.

Conclusion

The integration of artificial intelligence into the military kill chain represents a paradigm shift on par with the introduction of mechanization or nuclear weaponry. Across the F2T2EA cycle, AI has transitioned from a theoretical concept to a deployed reality. Systems like the Maven Smart System and The Gospel demonstrate that artificial intelligence is already revolutionizing the Find, Fix, and Target phases, processing insurmountable volumes of intelligence data to orchestrate battlespace command and control at a speed that dwarfs human capability. Simultaneously, the proliferation of terminal autonomy in munitions like the Lancet, Saker Scout, and those pursued under the Replicator Initiative proves that the Engage phase is rapidly moving beyond direct human control to counter the realities of electronic warfare. As global powers race to master CJADC2 and Intelligentized Warfare, the strategic landscape of the twenty-first century will be defined by algorithmic supremacy, sensor fusion, and distributed swarm consensus. However, the pursuit of machine-speed warfare carries immense and unprecedented risks. The tension between the desire to close the kill chain in seconds and the legal, ethical, and operational imperative to maintain meaningful human judgment remains fundamentally unresolved. Moving forward, the true test of military power will not solely be the technological capability to automate the kill chain, but the institutional capacity to ensure that these hyper-fast, highly autonomous systems remain robust against adversarial deception, strictly aligned with international humanitarian law, and permanently subservient to human strategic intent.

Works cited

1. C-UAS Kill Chain (F2T2EA) \- The Lightning Press SMARTbooks, https://www.thelightningpress.com/c-uas-kill-chain-f2t2ea/

2. Kill chain (military) \- Wikipedia, https://en.wikipedia.org/wiki/Kill\_chain\_(military)

3. AI Kill Chain: Autonomous Military Targeting \[2026\] \- Kunal Ganglani, https://www.kunalganglani.com/blog/ai-kill-chain-military-targeting

4. Targeting at Machine Speed: The Capabilities—and Limits—of Artificial Intelligence, https://mwi.westpoint.edu/targeting-at-machine-speed-the-capabilities-and-limits-of-artificial-intelligence/

5. Winning the Kill Chain Competition \- Mitchell Institute for Aerospace Studies, https://www.mitchellaerospacepower.org/app/uploads/2023/05/Scale\_Scope\_Speed\_Survivability\_-KillChain\_-Policy\_Paper\_40-New.pdf

6. Mapping Artificial Intelligence to the Naval Tactical Kill Chain, https://nps.edu/documents/10180/142489929/NEJ+Hybrid+Force+Issue\_Mapping+AI+to+The+Naval+Kill+Chain.pdf

7. Is the PLA Overestimating the Potential of Artificial Intelligence? \- NDU Press, https://ndupress.ndu.edu/Media/News/News-Article-View/Article/4042798/is-the-pla-overestimating-the-potential-of-artificial-intelligence/

8. Radar Automatic Target Recognition (RATR) \- Emergent Mind, https://www.emergentmind.com/topics/radar-automatic-target-recognition-ratr

9. A Comprehensive Survey on SAR ATR in Deep-Learning Era \- MDPI, https://www.mdpi.com/2072-4292/15/5/1454

10. Deep Learning-Based SAR Target Recognition: A Dual-Perspective Survey of Closed Set and Open Set \- MDPI, https://www.mdpi.com/2076-3417/15/23/12501

11. Review of Synthetic Aperture Radar Automatic Target Recognition: A Dual Perspective on Classical and Deep Learning Techniques \- IEEE Xplore, https://ieeexplore.ieee.org/iel8/4609443/4609444/11082447.pdf

12. Automatic Target Recognition on Synthetic Aperture Radar Imagery: A Survey | Request PDF \- ResearchGate, https://www.researchgate.net/publication/348734355\_Automatic\_Target\_Recognition\_on\_Synthetic\_Aperture\_Radar\_Imagery\_A\_Survey

13. \[1801.08558\] Deep Learning for End-to-End Automatic Target Recognition from Synthetic Aperture Radar Imagery \- arXiv, https://arxiv.org/abs/1801.08558

14. (PDF) Investigation of the Robustness and Transferability of Adversarial Patches in Multi-View Infrared Target Detection \- ResearchGate, https://www.researchgate.net/publication/397033315\_Investigation\_of\_the\_Robustness\_and\_Transferability\_of\_Adversarial\_Patches\_in\_Multi-View\_Infrared\_Target\_Detection

15. Exact and Heuristic Methods for the Weapon Target Assignment Problem \- DSpace@MIT, https://dspace.mit.edu/entities/publication/8c3b5dbd-451a-4b36-b895-39fedeb34e02

16. Exact and Heuristic Algorithms for the Weapon-Target Assignment Problem \- PubsOnLine, https://pubsonline.informs.org/doi/10.1287/opre.1070.0440

17. Decentralized Weapon-Target Assignment under Asynchronous Communications \- CORE Lab, https://corelab.mae.ufl.edu/papers/WTA.pdf

18. Weapon-Target Assignment Problem with Interference Constraints using Mixed-Integer Linear Programming \- arXiv, https://arxiv.org/pdf/1911.12567

19. Generalized Intelligence for Tactical Decision-Making: Large Language Model–Driven Dynamic Weapon Target Assignment \- arXiv, https://arxiv.org/html/2511.10207v1

20. Solving the Dynamic Weapon Target Assignment Problem by an Improved Multiobjective Particle Swarm Optimization Algorithm \- MDPI, https://www.mdpi.com/2076-3417/11/19/9254

21. Weapon Target Assignment with Combinatorial Optimization Techniques \- The Science and Information (SAI) Organization, https://thesai.org/Downloads/IJARAI/Volume2No7/Paper\_7-Weapon\_Target\_Assignment\_with\_Combinatorial\_Optimization\_Techniques.pdf

22. DiRAC \- Distributed Robot Awareness and Consensus \- arXiv, https://arxiv.org/html/2510.16850v1

23. \[2301.13276\] Distributed Swarm Intelligence \- arXiv, https://arxiv.org/abs/2301.13276

24. SwarmRaft: Leveraging Consensus for Robust Drone Swarm Coordination in GNSS-Degraded Environments \- arXiv, https://arxiv.org/html/2508.00622v2

25. Efficient Distributed Consensus Algorithm For Swarm Robotic | Request PDF, https://www.researchgate.net/publication/365271614\_Efficient\_Distributed\_Consensus\_Algorithm\_For\_Swarm\_Robotic

26. SwarnRaft: Leveraging Consensus for Robust Drone Swarm Coordination in GNSS-Degraded Environments \- arXiv, https://arxiv.org/html/2508.00622v1

27. \[2508.18153\] DANCeRS: A Distributed Algorithm for Negotiating Consensus in Robot Swarms with Gaussian Belief Propagation \- arXiv, https://arxiv.org/abs/2508.18153

28. \[2601.17303\] Decentralized Multi-Agent Swarms for Autonomous Grid Security in Industrial IoT: A Consensus-based Approach \- arXiv, https://arxiv.org/abs/2601.17303

29. Autonomous Multi-domain Adaptive Swarms-of-Swarms (AMASS) \- SAM.gov, https://sam.gov/opp/2929a61ea2bd44f7a68095449e1fd68d/view

30. Project Maven \- Wikipedia, https://en.wikipedia.org/wiki/Project\_Maven

31. Maven Smart System, https://assets.ctfassets.net/xrfr7uokpv1b/25muZs93DY5XOUBtuh68yn/b40d9784f69917219f972890ceede986/Maven\_One\_Pager.pdf

32. What Is Maven Smart System, and What Does It Do? \- CSIS, https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do

33. Palantir Expands Maven Smart System AI/ML Capabilities to Military Services, https://investors.palantir.com/news-details/2024/Palantir-Expands-Maven-Smart-System-AIML-Capabilities-to-Military-Services/

34. DOD components face 'aggressive' timeline for Maven Smart System transition, https://defensescoop.com/2026/04/15/palantir-maven-smart-system-pentagon-program-transition-feinberg/

35. Maven Smart System: Innovating for the Alliance \- Palantir Blog, https://blog.palantir.com/maven-smart-system-innovating-for-the-alliance-5ebc31709eea

36. Maven Smart System \- Missile Defense Advocacy Alliance, https://www.missiledefenseadvocacy.org/maven-smart-system/

37. Incident 672: 'Lavender' and 'The Gospel' AI Systems Reportedly Used in Gaza Targeting Operations with Civilian Harm Allegations, https://incidentdatabase.ai/cite/672/

38. AI-assisted targeting in the Gaza Strip \- Wikipedia, https://en.wikipedia.org/wiki/AI-assisted\_targeting\_in\_the\_Gaza\_Strip

39. Israel – Hamas 2024 Symposium \- The Gospel, Lavender, and the Law of Armed Conflict, https://lieber.westpoint.edu/gospel-lavender-law-armed-conflict/

40. Israel's use of Military AI-systems in the occupied territories of Palestine \- Overheid.nl, https://puc.overheid.nl/mrt/doc/PUC\_794610\_11/1/

41. Questions and Answers: Israeli Military's Use of Digital Tools in Gaza | Human Rights Watch, https://www.hrw.org/news/2024/09/10/questions-and-answers-israeli-militarys-use-of-digital-tools-in-gaza

42. Lavender, Where's Daddy and the ethics of AI-driven war \- Down To Earth, https://www.downtoearth.org.in/governance/lavender-wheres-daddy-and-the-ethics-of-ai-driven-war

43. Israel's AI-enabled targeting of Hamas members jeopardizes moral and legal standards of warfare \- Background \- Utrecht University, https://www.uu.nl/en/achtergrond/israels-ai-enabled-targeting-of-hamas-members-jeopardizes-moral-and-legal-standards-of-warfare

44. Emerging patterns in intelligentized warfare: speed, consumption, and the transformation of military operations \- Frontiers, https://www.frontiersin.org/journals/political-science/articles/10.3389/fpos.2026.1837867/full

45. How Russia Is Building a Sovereign Drone Ecosystem for AI-Driven Autonomy \- CSIS, https://www.csis.org/analysis/how-russia-building-sovereign-drone-ecosystem-ai-driven-autonomy

46. How Ukrainian drones use AI and 'autopilot' to outsmart Russian jamming \- Reddit, https://www.reddit.com/r/UkrainianConflict/comments/1uo1128/how\_ukrainian\_drones\_use\_ai\_and\_autopilot\_to/

47. Ukraine's Armed Forces will use SAKER SCOUT drones with AI | Ukrainska Pravda, https://www.pravda.com.ua/eng/news/2023/09/04/7418331/

48. Russia Probably Has Not Used AI-Enabled Weapons in Ukraine, but That Could Change \- CSIS, https://www.csis.org/analysis/russia-probably-has-not-used-ai-enabled-weapons-ukraine-could-change

49. Year Ahead – The U.S. DoD Replicator Initiative and the Acquisition Process for Autonomous Weapons \- Lieber Institute, https://lieber.westpoint.edu/us-dod-replicator-initiative-acquisition-process-autonomous-weapons/

50. Scaling the Future: How Replicator Aims to Fast-Track U.S. Defense Capabilities, https://warontherocks.com/scaling-the-future-how-replicator-aims-to-fast-track-u-s-defense-capabilities/

51. Implementing DoD Replicator Initiative at Speed and Scale \- Defense Innovation Unit, https://www.diu.mil/latest/implementing-the-department-of-defense-replicator-initiative-to-accelerate

52. DoD promised a 'swarm' of attack drones. We're still waiting. \- Responsible Statecraft, https://responsiblestatecraft.org/replicator/

53. https://www.belfercenter.org/replicator-autonomous-weapons-taiwan

54. Pentagon turns Replicator focus to C-sUAS capabilities \- Unmanned airspace, https://www.unmannedairspace.info/counter-uas-systems-and-policies/pentagon-turns-replicator-focus-to-c-suas-capabilities/

55. Pentagon Makes First Replicator 2 Counter-Drone Buy \- MeriTalk, https://www.meritalk.com/articles/pentagon-makes-first-replicator-2-counter-drone-buy/

56. Pentagon homes in on counter-drone tech in Replicator 2 initiative \- Breaking Defense, https://breakingdefense.com/2024/09/pentagon-homes-in-on-counter-drone-tech-in-replicator-2-initiative/

57. Replicator (United States military) \- Wikipedia, https://en.wikipedia.org/wiki/Replicator\_(United\_States\_military)

58. Replicator 2: Pentagon's New Plan to Combat Drone Attacks \- YouTube, https://www.youtube.com/watch?v=w7J26Xq4TZo

59. What USAF Is Doing to Defend Bases from Drone Attack \- Air & Space Forces Magazine, https://www.airandspaceforces.com/air-force-experts-strategies-defend-bases-drone-attack/

60. Joint Interagency Task Force Announces First Replicator 2 Purchase to Counter Homeland Drone Threats \- Department of War, https://www.war.gov/News/News-Stories/Article/Article/4377021/joint-interagency-task-force-announces-first-replicator-2-purchase-to-counter-h/

61. Exploring PRC Research on Kill Chains and Kill Webs: How Might it Help the PLA? | CNA, https://www.cna.org/analyses/2026/07/exploring-prc-research-on-kill-chains-and-kill-webs

62. China's Ambitions for AI-Driven Future Warfare \- CSBA, https://csbaonline.org/chinas-ambitions-for-ai-driven-future-warfare/

63. "AI weapons" in China's military innovation \- Brookings Institution, https://www.brookings.edu/articles/ai-weapons-in-chinas-military-innovation/

64. The People's Liberation Army's Perspectives on Artificial Intelligence Highlighting Integration as Key to "Intelligentization" Goals \- RAND Corporation, https://www.rand.org/pubs/perspectives/PEA4574-1.html

65. China's Military Employment of Artificial Intelligence and Its Security Implications, https://www.iar-gwu.org/print-archive/blog-post-title-four-xgtap

66. ARTIFICIAL INTELLIGENCE DoD Directive 3000.09: Autonomy in Weapon Systems \- Carahsoft, https://static.carahsoft.com/concrete/files/2417/3887/5530/Guidance\_DoD\_Directive\_3000.09\_-\_Autonomy\_in\_Weapon\_Systems.pdf

67. Department of Defense Directive 3000.09 \- Wikipedia, https://en.wikipedia.org/wiki/Department\_of\_Defense\_Directive\_3000.09

68. DoD Directive 3000.09, "Autonomy in Weapon Systems," January 25, 2023 \- Executive Services Directorate, https://www.esd.whs.mil/portals/54/documents/dd/issuances/dodd/300009p.pdf

69. United States, Use of Autonomous Weapons \- How does law protect in war? \- ICRC, https://casebook.icrc.org/case-study/united-states-use-of-autonomous-weapons

70. Background Briefing Review of the 2023 US Policy on Autonomy in Weapons Systems, https://humanrightsclinic.law.harvard.edu/wp-content/uploads/2023/02/Review-of-the-2023-US-Policy-on-Autonomy-in-Weapons-Systems.pdf

71. The Risks of Artificial Intelligence in Weapons Design | Harvard Medical School, https://hms.harvard.edu/news/risks-artificial-intelligence-weapons-design

72. Inside the Pentagon's AI Kill Chain \- YouTube, https://www.youtube.com/watch?v=T\_3ST-GOPEk

73. Anti-AI Countermeasures in Warfare: Terra Incognita for IHL? \- Lieber Institute \- West Point, https://lieber.westpoint.edu/anti-ai-countermeasures-warfare-terra-incognita-ihl/