Artificial intelligence already sits inside the modern kill chain. It does not yet replace the commander who authorises a strike, but it compresses the time between detection and engagement in ways human staffs cannot match when dozens of threats appear at once. The practical question for defence planners is no longer whether machines will take part in war. It is how far states will allow software to select, track, and, in a growing number of cases, engage targets before a person intervenes. Nowhere has that question been tested more directly, or more controversially, than in the opening months of 2026.
The Software Doing the Sorting: Maven and the New American Way of War
Most operational systems today keep a human in or on the loop, and the clearest example of that pattern is Palantir’s Maven Smart System, which Deputy Secretary of Defense Steve Feinberg formally designated a Pentagon program of record on March 9, 2026, a status that guarantees long-term funding and standardises the software’s adoption across every U.S. military service rather than leaving it as a collection of pilot projects and short-term contracts. Palantir is now seeking roughly $2.3 billion over five years to expand the system, which fuses satellite imagery, drone video, radar signals, and signals intelligence into a single common operating picture before ranking targets for a human officer to review. Pentagon officials describe the shift in blunt terms. Instead of consulting eight or nine separate systems to make a combat decision, Cameron Stanley, the Pentagon’s chief digital and AI officer, has explained that a commander can now identify a target and select a strike package for it simply by clicking on a screen. “We’ve gone from identifying the target to now coming up with a course of action to now actioning that target, all from one system,” Stanley told Palantir’s AIPCon 9 conference in March. “This is revolutionary.”
The empirical test of that claim arrived almost immediately. Operation Epic Fury, the U.S. air and missile campaign against Iran that began February 28, 2026, became the most intensive live demonstration of AI-enabled targeting any military has yet conducted. U.S. forces struck nearly 900 targets in the first twelve hours of the war and more than 13,000 over the following 38 days, according to White House and Central Command figures, a tempo officials have directly credited to Maven’s ability to process a data backlog that would previously have required thousands of human analysts. Retired Admiral Mark Montgomery told CBS News the military was processing roughly a thousand targets a day with a turnaround under four hours, a pace no previous American air campaign has matched. Officials have drawn an explicit historical comparison: identifying the same volume of targets during the 2003 invasion of Iraq required a team of roughly 2,000 intelligence analysts, work that Maven and a few dozen operators can now approximate in a fraction of the time. “Our war fighters are leveraging a variety of advanced AI tools,” said Admiral Brad Cooper, commander of U.S. Central Command, in a March briefing. “These systems help us sift through vast amounts of data in seconds so our leaders can cut through the noise and make smarter decisions faster than the enemy can react.”
That speed has come with a genuine cost attached. A U.S. cruise missile strike on the Shajareh Tayyebeh girls’ primary school in Minab, Iran, on the opening day of the war killed more than 170 people, most of them children, after the facility, once part of a military base but long since converted to civilian use, was added to an AI-generated target list without adequate human verification that the conversion had occurred. Central Command’s preliminary assessment attributed the error to outdated intelligence maps rather than a flaw in the software itself, but the incident has sharpened a concern researchers call automation bias, the tendency of human reviewers to rubber-stamp a machine’s recommendation rather than independently verify it once they have seen the system perform reliably thousands of times in succession. “There’s a concern that targeting approval could end up just being a mere formality because of the automation bias, where people are just relying on what the machine is telling them,” Nilza Amaral, head of research at Chatham House’s Global Governance and Security Centre, told The National. Future Military Technologies has examined this same tension in detail in a broader look at how the Pentagon is deploying AI, where the software nominates a course of action and a person still signs off on it, but the time available for that person to meaningfully scrutinise the recommendation keeps shrinking as the system’s throughput increases.
When the Software’s Maker Says No: The Anthropic Fight
One of the more consequential storylines of 2026 has not been a new weapon at all, but a dispute over who gets to set the rules for the AI already inside the kill chain. In late 2024, Anthropic’s Claude models were merged into the Maven architecture to provide enhanced targeting capability, and Claude briefly became the first frontier AI system cleared for the Pentagon’s most sensitive classified networks. That arrangement did not survive contact with the Trump administration’s demand that Claude be made available for what the Pentagon called “all lawful purposes,” a phrase Anthropic understood to include removing the usage restrictions it had built into its terms of service, restrictions that specifically barred using Claude for mass surveillance of American citizens or for developing fully autonomous weapons systems capable of engaging targets without human oversight. Anthropic declined, and in February 2026 Defense Secretary Pete Hegseth designated the company a supply chain risk, a rare classification that barred the Pentagon and its contractors from using Claude at all and triggered a Trump administration order for every federal agency, not only the Department of Defense, to phase the company’s technology out within six months.
Anthropic sued, and the legal outcome has been genuinely split. A federal judge in California ruled in August 2026 that the designation constituted unlawful retaliation against Anthropic for exercising its First Amendment rights and violated the company’s due process protections, calling the government’s action “arbitrary and capricious.” A separate case before the U.S. Court of Appeals for the District of Columbia Circuit went the opposite way. In a 2-1 decision issued September 25, 2026, the appeals court upheld the Pentagon’s designation, with the majority accepting Defense Undersecretary Emil Michael’s argument that the government need not rely on an AI supplier whose self-imposed restrictions could interfere with essential military operations. Michael celebrated the ruling in blunt terms on social media: “The hammer of justice has smashed @AnthropicAI arguments. They are a supply-chain risk to the defense industrial base… Warfighters will sleep better knowing that no private company will insert their opinions in the chain of command.” Anthropic has said it is considering further appeal, including a possible petition to the Supreme Court, while continuing to grow its commercial business roughly tenfold year over year ahead of a planned initial public offering.
The practical effect on the battlefield has been to replace one frontier model with others. OpenAI and additional AI developers have been recruited to assume the targeting role Claude previously played inside Maven, and the Pentagon’s broader May 2026 decision to bring eight technology companies, including Google, Microsoft, Amazon, NVIDIA, SpaceX, Oracle, and the startup Reflection AI, onto its classified networks reflects a deliberate strategy of redundancy: no single AI company should be able to unilaterally set the ethical terms under which its models are used for lethal purposes, because any company unwilling to meet the Pentagon’s usage requirements can, in principle, be replaced by one of several others holding equivalent contracts. Whatever the merits on either side of the underlying dispute, a matter still working through the American court system, the episode is the clearest evidence yet that the argument over how far AI should be trusted with targeting decisions is no longer a purely academic one. It is being fought, and decided, inside a courtroom in Washington, D.C.
Kinetic Effectors Still Fire Under Rules People Write
That same pattern, AI doing the sorting while a kinetic effector fires under rules of engagement written by people, repeats across the weapons already entering service. Lockheed Martin’s Sanctum battle manager has ingested radar tracks from a Fortem sensor, maintained a lock on a Shahed-style one-way attack drone, and cued a Joint Air-to-Ground Missile in a live-fire sequence the company assembled from concept to intercept in just 45 days, using existing, proven components rather than a new weapon developed from scratch. Northrop Grumman’s Raid Hunter, a 50mm gun-based air defence system built specifically to counter the economics of mass drone raids, uses AiON command software to correlate tracks and sequence engagements against mixed aerial threats, then hands the shot to a networked controller that can sit inside the Army’s Integrated Battle Command System, coordinating with Patriot, IRIS-T, and NASAMS batteries rather than operating as an isolated point-defence gun. The economics of swarm defence, cheap drones costing a fraction of the interceptors historically fired at them, are what is driving this entire category of architecture into existence, and Future Military Technologies has tracked that cost-exchange problem across multiple systems this year, from gun-based effectors to dedicated interceptor drones built to physically ram an incoming threat. In every one of these systems, AI’s job is speed and correlation. The decision to fire still rests with rules of engagement a human wrote in advance.
The Air Domain Moves Fastest: Loyal Wingmen Go From Paper to Production
Nowhere is machine teaming advancing faster than in the air, and the pace over the past several months has been genuinely remarkable. On June 17, 2026, the U.S. Air Force awarded production contracts to General Atomics and Anduril for the first increment of its Collaborative Combat Aircraft programme, four months ahead of schedule, converting what had been the prototype YFQ-42A Dark Merlin and YFQ-44A Fury into production aircraft redesignated FQ-42A and FQ-44A. The Air Force has committed to at least 150 of these aircraft by the end of the decade as a down payment toward a long-term goal of roughly 1,000 Collaborative Combat Aircraft, at a target unit cost near $30 million, about one-third the flyaway price of an F-35. Air Force Chief of Staff General David Allvin made the point plainly when describing just how quickly the programme moved from paper to hardware: “We have two prototypes of Collaborative Combat Aircraft that were on paper less than a couple of years ago.” Future Military Technologies has covered Anduril’s parallel push into autonomous rotorcraft with its Thunder programme, which applies the same crewed-uncrewed teaming logic to the low-altitude fight helicopters occupy, and Anduril’s Fury itself has already flown captive-carry tests with a live AIM-120 AMRAAM missile, with separation and live-fire trials planned before the end of 2026.
Türkiye’s Bayraktar Kızılelma offers the clearest evidence yet that this kind of autonomy is not confined to American programmes or to defensive missions. The 8.5-tonne unmanned combat aircraft, carrying a 1.5-tonne payload out to a combat radius near 500 nautical miles, performs its takeoff, landing, taxi, and cruise phases entirely autonomously, requiring two human operators only during the takeoff and landing sequence and a single operator once the aircraft reaches its mission phase. In late 2025, Kızılelma executed what Baykar has described as a world-first: identifying and striking an aerial target beyond visual range using its own Murad AESA radar and a Gökdoğan missile, entirely without a pilot in the loop for that specific engagement. Two prototypes have also flown a fully autonomous close-formation flight, and the aircraft is scheduled to begin operating from the Turkish Navy’s carrier-capable amphibious assault ship TCG Anadolu. Loyal wingmen do not yet decide national strategy. They already fly, and in at least one documented case fight, parts of the tactical mission that used to require a second cockpit.
The U.S. Air Force’s own HAVE HEAT trials, flown on the X-62 VISTA variable-stability testbed, showed an AI agent ingesting live infrared data from a Legion Pod, the same passive sensor system Future Military Technologies examined when Lockheed Martin networked it between two F-16s earlier this year, and flying a complete intercept geometry without a human hand on the stick, a demonstration that autonomous air-to-air engagement is no longer confined to simulation.
The Ground Picture: Where the Radio Link Ends and the Airframe’s Own Judgment Begins
On the ground, the picture is considerably more mixed, and the reason comes down to a hard technical constraint rather than a policy choice. First-person-view drones have become the principal tank killers in Ukraine, and a landmark RUSI report by Jack Watling and Nick Reynolds, Tactical Developments During the Third Year of the Russo-Ukrainian War, found that tactical unmanned aerial systems now account for 60 to 70 percent of damaged and destroyed Russian equipment, roughly twice the combined effect of every other weapon in the Ukrainian arsenal. Most of those drones still have a human pilot flying in first-person-view goggles. A growing minority now carry onboard machine vision capable of completing the final seconds of an attack independently once the radio link between pilot and drone drops or jamming blinds the operator, precisely the scenario Ukrainian and Russian electronic warfare units spend their days trying to engineer.
Ukraine’s Saker Scout, approved for service in 2023 and developed by a startup originally focused on agricultural crop-protection vision systems, can classify up to 64 categories of Russian military equipment, including vehicles under camouflage netting or tree cover, and continue an attack autonomously after losing its GPS signal, feeding its targeting data through Ukraine’s Delta battlefield-management system. Russia’s answer, according to detailed reporting from Ukraine’s military intelligence directorate, is the V2U loitering munition, which runs its target recognition on a Chinese-manufactured Leetop A203 minicomputer built around an Nvidia Jetson Orin processor, the same commercial-grade embedded computing chip sold openly to students and hobbyist developers worldwide. Ukrainian intelligence assesses that the V2U carries only a single GPS module, a design choice interpreted as an acknowledgment that Ukrainian electronic warfare has made satellite navigation unreliable near the front, and that the drone instead relies on computer vision, comparing its live camera feed against pre-loaded terrain imagery stored on an onboard solid-state drive, to navigate and identify targets once GPS and radio contact are lost. Those specific technical claims, drawn from a single intelligence disclosure, remain difficult to verify independently in full, yet the broader direction of travel is not in serious dispute among analysts on either side of the conflict: the last seconds of an attack are migrating from a human pilot’s judgement onto the airframe’s own onboard processor.
It is worth being precise about the limits of this shift, because the same RUSI report that documents drones’ dominant share of Russian losses is unusually blunt about how far autonomy still has to go before it approaches anything resembling reliability. Between 60 and 80 percent of Ukrainian FPV attempts fail to reach their intended target at all, depending on the sector of the front and the skill of the individual operator, and of those that do strike an armoured vehicle, a majority fail to destroy it outright. Weather grounds sorties. Electronic warfare blinds guidance links for extended stretches. Fibre-optic-guided drones, which trail a physical control cable specifically to defeat radio-frequency jamming, solve one problem while introducing new mechanical and logistical ones of their own. Future Military Technologies examined this same durability question in a recent look at how tanks are adapting to the drone age, and the underlying lesson applies equally to the drones themselves: AI meaningfully improves the probability that a given strike connects with its target. It does not repeal the physics of weather, jamming, or a determined active protection system, and it does not repeal the basic logistics of getting enough functioning drones to the front line in the first place.
Command Software: Compressing the Human Decision Cycle
If any single layer of this transformation deserves the label revolutionary, it is command and control software, because this is where the largest near-term change to how militaries actually operate is concentrated. The British Army’s ASGARD prototype compressed corps-level planning from 72 hours down to a single hour during NATO trials conducted in Estonia, according to Chief of the General Staff General Sir Roly Walker, allowing a corps to prosecute roughly ten times as many targets in a day as it could before. Anduril’s Lattice Mesh and the U.S. Army’s own Next Generation Command and Control prototype, validated at scale during the Army’s Project Convergence Capstone 5, aim to give every vehicle and soldier access to the same common operating picture, so that a drone spotting a target becomes a fire mission in seconds rather than the minutes or hours a traditional staff process would require.
That same logic is spreading across allied and industry programmes well beyond the United States and Britain. Rheinmetall’s Battlesuite, built on the German company blackned’s Tactical Core software, treats the battlefield as a plug-in application layer, in which sensors, radios, and unmanned systems from different manufacturers connect into a single shared data fabric rather than operating as isolated, incompatible stovepipes. Europe’s MARTE tank studies already assume that future armoured vehicles will operate permanently inside that kind of digital envelope, receiving targeting cues and threat warnings from a networked architecture rather than relying solely on their own onboard sensors. The consistent thread across every one of these programmes, British, American, and German alike, is the same: the decisive military advantage increasingly lives in the software fusing sensor data and generating a common picture, not in any single platform’s individual sensor or gun.
Policy Racing to Catch Up
Hardware and software have moved faster than the rules meant to govern them, and 2026 has been the year that gap became impossible to ignore. Department of Defense Directive 3000.09, first issued in 2012 and updated in 2023, still requires that autonomous and semi-autonomous weapon systems be designed to allow commanders and operators to exercise “appropriate levels of human judgment over the use of force.” On June 5, 2026, President Trump signed National Security Presidential Memorandum 11, ordering the Secretary of War to issue an update to that directive within 90 days, roughly September 3, 2026, and to review it annually thereafter to keep pace with what the memorandum calls the rapidly evolving capabilities of AI systems, while directing agencies to terminate contracts, to the maximum extent the law allows, with companies that repeatedly demonstrate conduct inconsistent with the memorandum’s push to accelerate AI adoption, language widely read as targeting exactly the kind of restrictions Anthropic had placed on Claude. Senator Ruben Gallego wrote to Secretary Hegseth on June 12, 2026, pressing for detail on what safeguards would survive a revision conducted on such a compressed timeline, warning that any update which “significantly reduces safeguards and disrupts rigorous reviews risks friendly fire and civilian” harm. Analysts at the Center for Strategic and International Studies have argued the directive’s core framework does not need an overhaul so much as sharper definitions, since some weapons already operate with a degree of autonomy in practice that their formal doctrinal classification does not fully capture.
The institutional and budgetary shift has been even more dramatic. The Pentagon’s Defense Autonomous Warfare Group, absorbed from the earlier Replicator initiative and housed within U.S. Special Operations Command, received a modest $225.9 million in its first year of operation in fiscal 2026. The Trump administration’s fiscal 2027 budget request proposed increasing that figure to $54.6 billion, a roughly 24,000 percent single-year jump that would have made DAWG’s budget line larger than the entire proposed budget of the U.S. Marine Corps. Retired General and former CIA Director David Petraeus called it the largest single commitment to autonomous warfare in history. Secretary Hegseth has separately announced plans to stand up a dedicated sub-unified command for autonomous warfare, comparable in institutional status to Joint Special Operations Command, though five months after the initial request the political reality looks considerably more modest than the headline figure suggested: the House Armed Services Committee, Senate Armed Services Committee, and House Appropriations Committee have each independently written the actual FY2027 appropriation down to roughly $1 billion, and the promised sub-unified command has not yet materialised as its own distinct organisation.
Internationally, the policy gap is even wider. The International Committee of the Red Cross continues to press for a legally binding international instrument prohibiting unpredictable autonomous weapons whose effects cannot be adequately controlled. The Convention on Certain Conventional Weapons Group of Governmental Experts has met repeatedly on the subject without producing a treaty. One hundred and sixty-four states backed a 2025 United Nations General Assembly resolution calling for formal negotiations toward binding rules, yet none of the major AI and autonomous weapons developers, the United States, China, and Russia chief among them, has joined any proposed ban, leaving the most capable militaries in the world to write their own rules largely without external constraint.
The Most Honest Picture in 2026
The most accurate way to describe where things actually stand is layered rather than binary. AI already fights parts of wars. It sorts imagery, ranks targets, flies air-to-air intercepts, guides the terminal seconds of a loitering munition, and sequences gun and missile fire against drone raids faster than any human staff could manage unaided. Humans still set the objectives, write the rules of engagement, accept the legal responsibility, and, in nearly every declared Western system, personally authorise the release of lethal force against a person. That division of labour is going to keep shifting under the pressure of sheer volume. A raid of 200 cheap airframes arriving at once leaves no realistic window for a staff officer to individually evaluate each track before a decision has to be made, and software will inevitably recommend, then increasingly execute, more of that defensive engagement on its own. Offensive autonomy against human targets remains politically and legally radioactive for most democracies, the Anthropic dispute and NSPM-11’s compressed 90-day review both being direct evidence of just how contested that line still is, even as individual battlefield systems in Ukraine and Russia already operate in the grey zone once a communications link is lost.
Decisions will be made faster
Wars of the next decade will be fought by humans who lean on machines that think faster than they do. The Maven-enabled tempo of Operation Epic Fury, the production line now turning out Collaborative Combat Aircraft on both sides of the Atlantic, and a Pentagon directive being rewritten on a 90-day clock all point toward the same underlying reality: the machines will not replace the political decision to go to war, and no serious military power has proposed that they should. But they will decide, more often with each passing year, which vehicle in a column dies first, which target in a saturated raid gets engaged and which does not, and how much time the human standing at the end of that chain actually has left to say no.
