Anatomy of an AI Kill Chain: New Report Exposes How Militaries Are Automating Life and Death Decisions

A landmark report released this week is forcing a reckoning with one of the most consequential technological developments of our era: the systematic automation of military targeting through artificial intelligence. “Anatomy of an AI Kill Chain,” a joint visual investigation by the transparency watchdog Airwars and the AI Now Institute, provides the most detailed public accounting to date of how AI now permeates virtually every stage of the modern military kill chain — from sensor data collection to target engagement — across active conflicts in Ukraine, Gaza, and Iran.

The report’s central finding is as stark as it is alarming. According to the investigation, only two of the six stages of the U.S. military’s kill chain now involve humans in the loop, with a third involving some degree of human oversight. The remaining stages are now fully automated. This means that the process by which the military identifies, tracks, and decides to engage targets — a process that traditionally required deliberate human judgment at each step — has been compressed to near-machine speed, with human oversight reduced to a fraction of what it once was.

What Is a Kill Chain?

The military “kill chain” refers to the systematic, multi-stage process of finding and engaging a target. The U.S. military formalizes this as the F2T2EA model: Find, Fix, Track, Target, Engage, and Assess. Each stage has historically required human intelligence analysts, commanders, and operators to make critical decisions with life-or-death consequences.

The AI Now Institute and Airwars break the chain into six stages in their report: sensor and data gathering, surveillance and pattern recognition, target identification and nomination, weapon and route planning, engagement execution, and battle damage assessment. What their investigation reveals is that AI systems have penetrated every single one of these stages — not as theoretical future capabilities, but as deployed, operational systems being used right now in live combat zones.

The kill chain is no longer a sequential, human-paced process. AI has collapsed the time from intelligence gathering to strike from days or weeks to, in some cases, mere minutes or seconds. Palantir’s Maven Smart System, for instance, has been used to identify thousands of targets in the Iran conflict, dramatically accelerating the pace of operations.

The Systems Behind the Automation

The report identifies a constellation of AI tools now embedded in military operations. Palantir’s Project Maven — the same system that sparked employee protests at Google in 2018 — has evolved into a full-spectrum targeting platform. It processes vast streams of surveillance data from satellites, drones, and ground sensors, using AI models to identify potential targets and recommend strikes.

Perhaps most controversially, the investigation documents the use of Anthropic’s Claude model in military contexts. Claude has reportedly been integrated into Israel’s targeting pipeline, feeding into systems like “Lavender” and “Gospel” — AI programs used to identify human targets and infrastructure. Lavender, in particular, was programmed to accept a threshold of up to 100 civilian casualties per target in some configurations, according to earlier reporting. That a model built by a company that publicly brands itself around “AI safety” is now embedded in the military kill chain represents a profound irony that the report does not let pass.

The Register’s coverage of the report notes that it maps “the risks of ubiquitous surveillance and flawed algorithms” — emphasizing that these systems operate on training data that is inherently biased, incomplete, and prone to producing false positives. When the cost of a false positive is a human life, the stakes of algorithmic error become existential.

The Three Conflicts: Ukraine, Gaza, and Iran

The report draws on operational data from three concurrent conflicts that, taken together, represent the first large-scale deployment of AI in warfare.

In Ukraine, the conflict has demonstrated how cheap, expendable autonomous hardware — particularly drones — can spread at scale. AI-powered drones now handle their own target acquisition and terminal guidance, reducing the role of human operators to little more than authorization at the launch point. The sheer volume of autonomous systems deployed has overwhelmed traditional air defense and created a new paradigm of swarm warfare.

In Gaza, AI has moved into what researchers describe as the “cognitive core” of military targeting. Systems like Lavender and Gospel process signals intelligence, communications intercepts, and behavioral patterns to generate target lists at a pace that human analysts cannot match — or meaningfully review. The result has been a dramatic expansion of the target universe, with AI systems nominating individuals for strikes based on algorithmic assessments of their suspected affiliations.

In Iran, Palantir’s Maven Smart System was deployed at unprecedented scale. The U.S. military reported striking over 11,000 targets since the beginning of the conflict, a tempo that would have been impossible without AI-assisted targeting. The system compressed the intelligence-to-strike cycle to a degree that military planners themselves have described as transformative — and, according to critics, dangerously opaque.

The Erosion of Human Oversight

Dr. Heidy Khlaaf, the Chief AI Scientist at the AI Now Institute and the report’s lead author, has emerged as one of the most prominent voices warning about the militarization of AI. In an interview on Democracy Now! on August 13, she described the kill chain as “a series of different types of AI algorithms that are chained together, each with their own pitfalls, each with their own flaws.”

The core problem, Khlaaf argues, is not any single AI system but the compounding effect of chaining multiple imperfect systems together. Each stage introduces its own error rate. When those errors cascade through the chain — from data collection, through pattern recognition, to target nomination and engagement — the result is a system where no single human can meaningfully verify the chain of decisions that led to a strike. The illusion of human oversight persists, but the reality is that humans are increasingly rubber-stamping algorithmic outputs they have neither the time nor the information to question.

This is compounded by what the report calls “automation bias” — the documented tendency of human operators to trust automated recommendations even when they are wrong. When an AI system flags a building as a military target, the cognitive burden required to override that recommendation is substantial, and the institutional pressure to accept it is immense, particularly under the time constraints of active combat.

The Accountability Gap

One of the report’s most damning findings is the accountability vacuum that AI weapons systems create. When a targeting decision goes wrong — when civilians are killed based on an algorithmic error — there is no clear chain of responsibility. Is it the fault of the AI system’s developers? The military commanders who authorized the strike? The intelligence analysts who fed data into the system? The report argues that the diffusion of decision-making across humans and algorithms makes accountability nearly impossible, creating what legal scholars call a “responsibility gap.”

Airwars documented the first confirmed civilian killed in an AI-assisted strike earlier in 2026, marking a grim milestone. The victim’s death illustrates the real-world consequences of algorithmic warfare — consequences that are likely to multiply as these systems proliferate.

The Commercial AI Connection

The report draws a direct line between the commercial AI industry and military applications. Companies like Palantir, Google, and Anthropic have become deeply embedded in the military-industrial complex, providing the foundational models, data infrastructure, and analytical tools that power the automated kill chain. This relationship, the report argues, creates a fundamental conflict of interest for companies that simultaneously market consumer AI products and profit from military contracts.

Khlaaf has been particularly critical of Anthropic, noting that despite the company’s safety-first branding, its Claude model has been integrated into military targeting pipelines. The gap between corporate safety rhetoric and battlefield reality, she argues, represents one of the most dangerous developments in AI governance.

What Comes Next

The report calls for urgent regulatory action, including mandatory transparency requirements for AI systems used in military contexts, independent auditing of targeting algorithms, and the establishment of clear legal frameworks for accountability when AI-assisted strikes cause civilian harm. It also urges a moratorium on fully autonomous weapons systems — those capable of selecting and engaging targets without meaningful human intervention.

As conflicts in Ukraine, Gaza, and Iran continue to serve as live testing grounds for AI warfare, the window for establishing meaningful guardrails is narrowing. The “Anatomy of an AI Kill Chain” report represents the most comprehensive effort to date to document what is happening — and to demand that the public, policymakers, and the AI industry confront the consequences of automating the most irreversible decision humans can make.

The question is no longer whether AI will be used in war. It already is. The question is whether anyone will be held accountable when it gets it wrong.