From Project Maven to Operation Sindoor: AI’s new role in war

In an earlier age of warfare, the most powerful machine on the battlefield was the aircraft, the missile, the tank, or the aircraft carrier. Today, the decisive machine may be something less visible: a software layer that absorbs satellite imagery, drone feeds, intercepted signals, battlefield reports, and cyber logs and turns them into decisions.

The recent US-Iran war has given the world one of the clearest glimpses yet of this transformation. The US used AI services, alongside B-2 bombers and Tomahawk cruise missiles. While AI models may not have chosen Ayatollah Ali Khamenei as the target, they may have helped identify his precise location. Because, let us face it: AI has moved from the margins of military experimentation into the operational stack. The question is where it sits inside the chain that connects sensors to shooters, and how much power it has to shape what human commanders eventually decide.

This is the “kill chain”: find, fix, track, target, engage, and assess. It is how a state identifies something it wants to hit, confirms the target, authorises action, and then checks whether the strike worked. The weak point in this chain was not always the missile or the aircraft. It was the human ability to process overwhelming volumes of information in real time.

A modern operation can generate more data than a large newsroom, a bank, and a weather agency combined. One must consider satellite imagery, full-motion drone transmitted video, electronic intelligence systems (ELINT), network logs from cyber teams, reports from human sources, and open-source monitoring of social media, shipping, aircraft movements, and local media (collectively called Open Source Intelligence, or OSINT in short). The work of intelligence is to connect them before the opportunity disappears.

This is where AI enters.

The value of LLM

The first generation of modern battlefield AI was computer vision. The Pentagon’s Project Maven, launched in 2017, used machine learning to analyse imagery and video from endless drone footage. Its early purpose was narrow but transformative: to detect objects of interest, classify them, and present them to analysts faster. Maven evolved into something closer to a battlefield operating system. In March 2026, the Pentagon decided to make Palantir’s Maven AI system a core military command-and-control platform, turning it into a long-term “program of record”. It was able to analyse large data streams from satellites, drones, radar, and sensors to identify threats and had already supported thousands of US strikes, including the ones on Iran in recent weeks. AI is increasingly fusing the battlefield.

In a traditional targeting cycle, analysts may have to move between separate systems: imagery databases, signals intelligence reports, watchlists, maps, strike logs, legal reviews, and command dashboards. An AI-enabled system brings these into one shared operational picture. A suspected missile launcher is not merely a shape on an image; it becomes an entity tied to coordinates, previous sightings, possible unit affiliation, nearby air defences, historical movement, communications patterns, and potential collateral risks.

This is also where large language models (LLMs) such as Claude become important. They are not needed to spot a truck in a satellite image; computer vision systems do that better. Their value lies in reasoning across messy information. They can summarise long reports, translate material, compare conflicting assessments, retrieve relevant documents, draft target folders, flag missing evidence, and help analysts ask better questions of large databases. Anthropic has publicly said that Claude is deployed across US national security agencies for mission-critical applications, including intelligence analysis, modelling and simulation, operational planning, and cyber operations.

The likely architecture is layered. At the bottom are sensors: satellites, drones, radar, signals intelligence systems, cyber tools, and human intelligence reports. Above that sits a data-fusion layer, where platforms such as Maven align information into a common picture. Above that are specialised AI models: computer vision for detection, graph analytics for networks, anomaly detection for patterns, and language models for summarisation and reasoning. Above them are human analysts, lawyers, and commanders.

Why the Iran war is significant

That final human layer is still crucial. The autonomy directive of the Department of Defense (DoD) continues to require appropriate levels of human judgment over the use of force. But “human-in-the-loop” can mean different things. If an AI system narrows 10,000 possibilities to 20, ranks them by confidence levels, drafts the justification, and recommends weapons pairing, the human still decides, but the machine has shaped the menu.

That is why the Iran case is so significant.

The public evidence does not show that an AI model selected Khamenei or generated the strike decision. But AI-enabled reconnaissance and targeting support would have been highly relevant to finding leadership movements, mapping command networks, identifying military facilities, synchronising cyber and kinetic operations, and rapidly assessing damage after each strike. AI is NOT an assassin. It is a war room analyst with access to every file cabinet.

A human analyst may ask: “What facilities connected to Iran’s missile command have shown unusual activity in the last 48 hours?” The system may retrieve satellite changes, radar emissions, movement logs, known unit associations, and prior human intelligence reports and then produce a summary, such as, “three locations show abnormal activity, two are near civilian infrastructure, one has prior association with a specific command unit, and confidence is limited by lack of recent thermal imagery”.

That is not a kill decision. But it is a powerful step towards one.

AI systems testing possible courses of action

The technical term is retrieval-augmented generation, or RAG. It is like a bank chatbot answering from its training documents. In a military system, the same concept can be far more serious. The model does not “remember” the battlefield from training data, but it retrieves relevant classified data and reasons over that material. In effect, the model becomes a conversational interface over a classified intelligence estate.

Graph analysis is another. Leadership targeting is rarely about a single person. It is about networks which include aides, convoys, safe houses, communications nodes, family compounds, security units, religious offices, aircraft, vehicles, and recurring routes. It is like a social networking site that suggests possibly related users and moots an “Add Friend” option.

The next layer is simulation. The Pentagon’s 2026 AI strategy explicitly discusses agent networks, battle management, and decision support, from campaign planning to kill-chain execution. This means AI systems are not just helping identify targets; they are also increasingly being used to test possible courses of action. What happens if a radar site is hit first? How might Iran’s missile units respond? Which air-defence gaps open for 20 minutes? What escalation paths follow the killing of a senior leader? The model may not answer these questions alone, but it can help planners run more branches and sequels than a human staff could do manually.

Cyber operations add another dimension. In May 2026, the Pentagon was deploying the now controversial Anthropic’s Mythos model under “Project Glasswing” to detect and fix long-standing software vulnerabilities across US government systems. Cyber tools are often used to map networks, disrupt communications, monitor retaliation risk, harden US systems, or detect malware from rogue actors. AI can triage vulnerabilities, generate patches, analyse malware, and translate technical telemetry into command-level risk. The state that can combine cyber reconnaissance with physical strikes gains a deeper view of the adversary’s nervous system.

This is also where the ethics of speed become unavoidable. The same systems that can accelerate target discovery can also accelerate error. Old imagery, faulty geolocation, a mistranslated intercept, a misclassified building, or a pattern-of-life model trained on incomplete data can all converge into a wrong but confident answer.

The Minab school strike in Iran is a stark civilian-harm case study. One must not claim that AI caused the strike; there is no public evidence for that. But it is exactly the type of incident that shows why AI-compressed targeting raises hard questions.

US President Donald Trump addressed the fatal February 28 strike on a girls’ school in Minab, southern Iran, and said nobody attacked it intentionally, while an initial US military probe suggested US forces were likely responsible and indicated outdated intelligence. The school was near an Islamic Revolutionary Guard Corps–linked military site, a classic targeting dilemma in which civilian proximity, stale intelligence, and confidence scoring become matters of life and death.

Future of AI warfare

The future of AI warfare will be judged not only by whether it helps militaries strike faster but by whether it helps them avoid striking the wrong thing. A responsible AI targeting workflow should be able to surface civilian objects, show when the last reliable image was taken, flag protected sites such as schools and hospitals, expose uncertainty instead of hiding it, and force a pause when the evidence is thin. If the system only optimises speed, it may make war more efficient without making it more lawful or more humane.

AI-assisted targeting is not the same as an autonomous weapon. In the first, AI helps humans analyse, prioritise, and plan. In the second, a system can identify, select, and engage a target without direct human intervention. The line between the two can blur.

A commander may technically remain in the loop, but if the system presents only ranked targets, hides uncertainty, penalises delay, and frames non-action as operational failure, the human can become a rubber stamp. That is why the deeper question is not whether a human clicked approve. It is whether the human had enough time, context, authority, and dissenting evidence to make a real decision.

This is also why the competition between Anthropic and OpenAI is not simply Silicon Valley rivalry. It is about who goes to the hilt to bag the elusive contracts even as ethics and surveillance-related concerns surface. While Anthropic had an early lead, securing a $200 million DoD agreement, it also clashed with the Pentagon over safeguards concerning fully autonomous weapons and mass domestic surveillance. This paved the way for OpenAI’s entry.

India should be watching this closely.

A new normal

Operation Sindoor offered India its own glimpse of a faster, more data-driven form of conflict. The 88-hour campaign marked a “new normal” in counter-terror strategy, combining precision strikes, advanced technology, and coordinated action across domains. Recent commentary has also linked Operation Sindoor to improved situational awareness, integrated systems, and real-time intelligence sharing, exactly the areas that define modern AI-enabled command.

Union Defence Minister Rajnath Singh said India had halted the operation on its own terms and referred to the Sudarshan air defence system as a key example of AI in modern warfare. Reports on Akashteer, India’s automated air-defence control and reporting system, said it helped create an integrated air picture and supported the interception of drones and missiles during the conflict.

The lesson is not that Operation Sindoor was a fully AI-run campaign. That would overstate what is known. The lesson is more specific: India’s warfare has begun moving from being platform-centric to network-centric. Counter-drone defence, layered air defence, real-time air-picture sharing, precision munitions, cyber alerting, and strategic messaging— all became part of the same conflict cycle. The challenge is to make that cycle faster, more integrated, and more sovereign without allowing automation to outrun judgment.

The budget question is central here. India is not starting from zero, but its spending pattern still reveals a readiness gap. India’s FY2025-26 defence budget was Rs.6.81 lakh crore, with a large share absorbed by manpower costs, and around Rs.1.80 lakh crore earmarked for modernisation and procurement. This imbalance matters because AI-enabled warfare does not require only aircraft and missiles; it requires compute, classified cloud or air-gapped infrastructure, secure data lakes, sensors, software-defined command systems, cyber ranges, model evaluation labs, data labelling pipelines, and specialised personnel.

While the Ministry of Defence fully utilised its FY2025-26 capital outlay, full utilisation is not the same as transformation. India must ask whether enough of the capital and research-and-development spend is being directed towards the less glamorous but decisive layer of future war: intelligence, surveillance, and reconnaissance fusion, AI-enabled command-and-control, autonomous and semi-autonomous platforms, electronic warfare, secure chips, military-grade compute, and sovereign foundational models.

However, things are slowly moving. India’s indigenous air-defence push, including Project Kusha, has been described by Rajnath Singh as strategically important. While the IndiaAI Mission has invited proposals to build foundational AI models for domestic AI Infrastructure, India needs more for military AI sovereignty. A defence model must operate with classified data, in Indian-controlled environments, under Indian doctrine, with Indian audit trails and without dependence on a foreign provider that could change access, pricing, policy, or political terms during a crisis.

This is why Claude’s dispute with the US government is relevant to India. If even the Pentagon can find itself negotiating with a private AI company over red lines, model access, and permitted uses, India cannot assume that foreign frontier models will be available on Indian terms in wartime. AI sovereignty is much more than a Hindi or Indian language RAG-based chatbot.

Immediate implications for India

India’s gaps are therefore not only technological; they are organisational. Future AI warfare will require joint data infrastructure across the Army, Air Force, Navy, intelligence agencies, cyber command structures, and space assets. The harder problem is not buying another sensor. It is ensuring that the sensor’s data can be trusted, fused, queried, protected, and acted upon by the right commander at the right time.

AI-enabled operations also require commanders who understand confidence intervals, model uncertainty, data freshness, adversarial deception, and automation bias; lawyers who can interrogate an AI-generated target file; operators who know when a recommendation is strong; and procurement officials who can distinguish between a demo and a deployable classified system. In future wars, algorithmic literacy will be a combat skill.

For India, this has three immediate implications.

First, India needs sovereign military AI models and secure deployment environments. India’s Project Ekam and indigenous defence AI efforts point in this direction.

Second, India must focus less on isolated AI tools and more on fusion platforms. The lesson of Maven is key here.

Third, India must prepare for AI-enabled escalation management. Operation Sindoor showed that modern conflict can move quickly from punitive strikes to air defence, cyber activity, drone threats, and strategic messaging.

All these make human control more important. India must avoid two extremes. One extreme is rejecting military AI because of ethical concerns, as adversaries will not wait. The other is blindly automating command decisions in pursuit of speed. The correct approach is decision superiority with decision sovereignty.

The most profound change

Given China has invested heavily in intelligentised warfare, unmanned systems, AI-enabled command, and information dominance, in a Himalayan contingency, the decisive problem may not be only the number of troops or aircraft. It may be which side can fuse satellite imagery, unmanned aerial vehicle feeds, electronic intelligence, logistics, weather, terrain, and cyber information into usable decisions in real time. A mere delay of 30 minutes in identifying a build-up may matter more than a marginal difference in platform capability.

The most profound change, however, is psychological. AI changes the tempo of war. It encourages commanders to expect instant answers. It rewards the side with cleaner data. It punishes bureaucratic delay. All this must happen without hallucinations.

That is why transparency, testing, and doctrine matter. Militaries must know when to trust the machine, when to question it, and when to slow down. The most dangerous AI system is not one that openly says it is uncertain. It is one that gives a crisp answer to an unclear question.

The battlefield is no longer only land, sea, airspace, and cyber. It is also computation. And in that domain, the winner may not simply be the side with the most powerful weapon. It may be the side whose machines help its humans think faster while keeping them responsible enough to know when not to strike.

Saikiran Kannan is an APAC-based independent journalist and an AI strategist.

Also Read | The mega AI bubble

Also Read | India built a software economy on borrowed hardware. AI changes the equation