Ask most people what “AI in defense” means and they will describe a single scene: an autonomous drone finding and striking a target with no human involved. That scene captures a real and important part of what is happening, but it is a small part. The more accurate picture, as of late 2026, is of at least five distinct transformations happening at once, at different speeds, governed by different rules, and facing different obstacles: frontier AI language models being wired directly into classified military networks, sensor-to-shooter kill chains being compressed from days to minutes, computer vision doing the work of thousands of imagery analysts, maintenance crews predicting equipment failures before they happen, and a policy apparatus visibly straining to keep pace with all of it.
This piece works through each of those threads in turn, with specific programmes, contracts, and dates, before pulling them together into a single picture of where military AI actually stands today, as distinct from where the popular imagination places it.
Frontier AI Models Enter the Classified Network
The most structurally significant AI development in defense this year had nothing to do with a weapon system. On May 1, 2026, the Pentagon announced that it had finalised agreements allowing eight technology companies, OpenAI, Google, Microsoft, Amazon Web Services, NVIDIA, SpaceX, Oracle, and the startup Reflection AI, to make their AI systems available on the Department of Defense’s classified networks, up to Impact Level 6 and Impact Level 7, the classification tiers covering Secret and Top Secret information. The arrangement lets Pentagon officials draw on Google’s Gemini models and Distributed Cloud infrastructure, OpenAI’s generative models, Microsoft’s Azure Government cloud, Amazon’s Bedrock and GovCloud environments, NVIDIA’s inference hardware, SpaceX’s Starshield connectivity layer alongside its sister company xAI’s Grok models, and Oracle’s cloud infrastructure, all for what the Pentagon has described as “lawful operational use” spanning warfighting, logistics, intelligence, and enterprise administration.
The path to that eight-company announcement began a year earlier and involved a dispute worth understanding in its own right, because it illustrates a governance tension that runs through the entire military AI landscape. In July 2025, the Pentagon’s Chief Digital and Artificial Intelligence Office (CDAO) awarded parallel contracts worth up to $200 million each to four AI developers, Google, OpenAI, Anthropic, and xAI, marking the first time frontier AI labs had been brought onto U.S. defence contracts at this scale. Anthropic’s Claude models briefly became the first frontier AI system cleared for the Pentagon’s most sensitive networks. That arrangement did not hold. According to reporting from Axios and multiple other outlets, the Pentagon subsequently pressed Anthropic to make Claude available for “all lawful purposes” without the usage restrictions Anthropic had built into its terms, restrictions that reportedly included limits on using the model for mass surveillance of American citizens and for developing fully autonomous weapons systems capable of operating without human oversight. Anthropic did not accept those terms, and when the Pentagon finalised its expanded eight-company classified-network arrangement in May 2026, Anthropic was excluded by name, with several outlets reporting the company subsequently pursued legal action alleging retaliation. Reflection AI, a comparatively young lab founded by former DeepMind researchers and valued at roughly $5 billion in early 2026, was included in the eight-company group instead, a detail several industry observers read as a signal that the Pentagon wants a wider and more easily substitutable bench of AI suppliers rather than depending on a small number of labs.
Whatever the merits on either side of that specific dispute, a matter of ongoing legal proceedings this article does not attempt to adjudicate, the episode reveals something structurally important about how the Pentagon now approaches frontier AI: as Emil Michael, the Department’s Chief Technology Officer, put it plainly, “I need redundancy.” The resulting eight-vendor architecture is deliberately built so that no single AI company can unilaterally set the ethical terms under which its models are used for military purposes, since any company unwilling to meet the Pentagon’s usage requirements can, in principle, be replaced by one of several others holding equivalent contracts. That is a genuinely new dynamic in military technology procurement, where dependence on a sole-source supplier has traditionally been the norm rather than the exception, and it is one that will likely define how every future frontier AI lab approaches the question of whether, and on what terms, to work with defence customers.
Compressing the Kill Chain
Set beside that institutional drama, the operational use of AI to accelerate targeting and engagement decisions is a more mature and, in relative terms, less contested story, one Future Military Technologies has tracked closely across several individual programmes this year.
The British Army’s ASGARD system, covered in depth in July, compressed corps-level planning cycles from 72 hours to a single hour and allowed a single corps headquarters to prosecute roughly ten times as many targets in a day, according to Chief of the General Staff General Sir Roly Walker. The system works by fusing data from multiple sensors and software layers, Anduril’s Lattice, Helsing’s Altra targeting software, and a networked command layer, into a single AI-curated decision loop that surfaces validated targets to a human commander rather than requiring staff officers to manually correlate each piece of incoming sensor data by hand. The American equivalent, CJADC2 (Combined Joint All-Domain Command and Control), pursues an identical objective at a larger scale, demonstrated during the U.S. Army’s Project Convergence Capstone 5, where officials described the central technical challenge not as a shortage of data but as data trapped in incompatible formats across disconnected systems, precisely the kind of correlation and translation problem AI-driven common operating pictures are designed to solve.
At the level of an individual weapon system, Lockheed Martin’s Sanctum-Grizzly-JAGM counter-drone kill chain illustrates the same principle in miniature: an AI battle manager ingesting radar tracks, assessing threat priority, and cueing a missile launch, compressing a detection-to-engagement sequence that would once have depended on a human operator manually tracking a screen and making a judgment call under serious time pressure. Northrop Grumman’s newly unveiled Raid Hunter gun-based air defence system, covered here in August, follows an identical template of AI-networked battle management layered onto a mature physical weapon.
The consistent analytical thread across all of these programmes is that the artificial intelligence itself is rarely the headline component. It is Fortem’s radar, or Northrop’s chain gun, or Helsing’s targeting software, that a system actually uses to detect and engage a threat. What the AI battle-management layer contributes is speed: correlating multiple sensor feeds, ranking threats, and generating a recommended engagement faster than a human staff could manage manually, particularly when the number of simultaneous threats, a mass drone raid, for instance, exceeds what human working memory can track in real time.
AI at the Sensor: Seeing and Deciding at Machine Speed
A less publicised but equally consequential thread involves AI operating directly on raw sensor data, in some cases inside the cockpit or airframe itself, rather than at a headquarters level.
The U.S. Air Force’s HAVE HEAT test programme, flown from Edwards Air Force Base in April 2026 using the X-62 VISTA variable-stability testbed aircraft, demonstrated AI agents ingesting live infrared sensor data from Lockheed Martin’s Legion Pod, an infrared search-and-track system, and using that data to direct the aircraft to autonomously intercept an airborne target in real time, entirely without a human manually flying an intercept geometry. That test sits within a broader Air Force effort to validate autonomous behaviours before they are trusted with genuinely operational missions, and it reflects the same underlying computer-vision and sensor-fusion techniques that give systems like Sanctum and Fortem’s R40 radar the ability to distinguish a genuine threat from background clutter fast enough to matter.
This category of AI use, perception and split-second manoeuvre decisions made onboard a platform rather than at a command centre, is arguably the most technically demanding form of military AI currently being tested, precisely because there is no opportunity for a human to catch an error before it has already shaped the aircraft’s flight path. It is also the category where the testing and validation challenges discussed later in this piece are most acute.
AI for Sustainment: Predicting Failure Before It Happens
If AI-enabled targeting captures the public imagination, AI-enabled maintenance is where the technology is arguably delivering the most consistent, least controversial value today, precisely because it operates well away from any weapons-release decision.
The U.S. Air Force’s Predictive Analytics and Decision Assistant (PANDA) programme uses machine learning to monitor operational and sensor data across more than 3,000 aircraft spanning 16 platforms, flagging hundreds of predicted component failures before they occur and allowing maintainers to schedule repairs during planned downtime rather than reacting to an unexpected grounding. The Navy has pursued a parallel path, deploying AI-driven predictive maintenance aboard the destroyer USS Fitzgerald that, according to Navy reporting, flagged an at-risk component for replacement before it failed and potentially left the ship stranded, and separately contracting for hull-inspection robots that generate roughly 4.2 million data points per vessel inspected, compared with a few hundred data points collected through traditional manual inspection methods.
The scale of the underlying problem explains why this investment matters so much. The Department of Defense spends approximately $90 billion annually maintaining ground systems, ships, and aircraft, according to a U.S. Government Accountability Office review, and the same review found that despite DoD issuing interim predictive-maintenance policy guidance as far back as 2002, the military services made only limited practical progress until quite recently. Officials in the Air Force’s Rapid Sustainment Office have described the goal in straightforward terms: making use of “historical maintenance data or onboard sensor data, telemetry data, really anything that we have that’s useful,” to build the evidentiary case for performing maintenance at exactly the right moment rather than on a fixed calendar schedule that is often either too conservative or too late.
It is worth noting that AI-driven sustainment and the kind of manufacturing ingenuity Future Military Technologies covered in the case of a grounded C-17, whose damaged nose panel was ultimately fabricated using a robotic Incremental Sheet Forming process after no commercial vendor would produce a one-off replacement, are frequently discussed by Air Force officials as complementary parts of the same broader sustainment challenge. DefenseScoop’s own August 2026 reporting on the Air Force’s predictive-maintenance efforts specifically cited that same C-17, tail number 0194, as an illustrative case of the sustainment pressures driving the service toward every available advanced technology, predictive and generative alike, even as officials caution that most AI sustainment tools remain in early experimentation or market research rather than widespread operational use.
The Governance Problem: Policy Racing to Catch Up
Every trend above raises the same underlying question: who is actually deciding when an AI-assisted or AI-driven system engages a target, and what rules govern that decision. The formal American answer is Department of Defense Directive 3000.09, “Autonomy in Weapon Systems,” first issued in November 2012 and most recently updated in January 2023. Its central requirement is 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,” a deliberately flexible standard that does not mandate a human review every individual engagement but does require that human judgment remain meaningfully present somewhere in the system’s design and employment.
The Pentagon’s CDAO has built out a Responsible AI Toolkit to help programme offices demonstrate compliance with 3000.09’s requirements in practice, an effort that Matthew Johnson, the CDAO’s acting Responsible AI chief, has described as necessary precisely because the directive’s process, while mandatory, does not spell out in granular detail how a given programme should prove it meets the standard. That ambiguity has drawn direct criticism from Congress. In Senate testimony in May 2026, Senator Joni Ernst argued that Pentagon policy is not keeping pace with the reality of how AI-driven targeting is now being integrated with autonomous munitions at a speed and scale the original 2012 directive, and even its 2023 update, were not designed to contemplate. Pentagon officials have signalled they are considering a further update to 3000.09, though analysts at the Center for Strategic and International Studies reported in May 2026 that Michael Horowitz, the Defense Department’s Emerging Capabilities Policy Office director responsible for the directive, has suggested any revision is likely to preserve the existing approach rather than overhaul it, on the grounds that the fundamental framework “laid out a very responsible approach to the incorporation of autonomy and weapons systems.”
That tension, a technology moving in months, a governing directive built for a slower era, updated only twice in fourteen years, is not unique to the United States, and it echoes the broader international vacuum around autonomous weapons and space-based systems that Future Military Technologies has examined elsewhere. It is also, in a sense, the same tension visible in the Pentagon’s Anthropic dispute described above: different actors within the same broad AI-in-defence ecosystem disagreeing, sometimes sharply, about where exactly the line for “appropriate human judgment” and “lawful use” should sit, with no single authoritative referee positioned to resolve the disagreement quickly.
The Global Dimension
None of this is unfolding in a vacuum defined solely by American policy debates. China has pursued its own parallel push to integrate AI into targeting, logistics, and command systems, with defence-industrial displays, such as Norinco’s AI-coordinated brigade of armoured vehicles and drones shown at the 2024 Zhuhai Airshow, signalling an intent to match or exceed the pace of Western AI-enabled battle management. Russia’s own more improvised adaptations, including its recent reliance on scarce, high-value crewed platforms like the A-50U and Su-57 to counter Ukrainian drone raids, stand in instructive contrast to the AI-curated, software-first approach the U.S. and UK are pursuing, illustrating that the ability to build and field this kind of software architecture, not just the ambition to do so, is itself becoming a meaningful axis of military competition. The details of that broader U.S.-China AI competition, spanning chips, model capability, and defence integration, are numerous enough to merit their own dedicated treatment, but the essential point for this piece is that no major military power currently regards AI adoption as optional.
Talent, Industry, and a Changing Vendor Base
The institutional shape of the defence AI market has changed as quickly as the technology itself. Companies such as Anduril, Palantir, and Helsing, all founded well after the traditional prime contractors, have become central suppliers of the software layer, Lattice, the Maven Smart System’s targeting workflows, Altra, that increasingly determines how effective a given hardware platform actually is in practice. Traditional primes including Lockheed Martin and Northrop Grumman have responded by building their own comparable software, Sanctum and AiON respectively, rather than ceding that layer entirely to newer entrants. Meanwhile, the same frontier AI labs now contracted to the Pentagon are simultaneously the most sought-after employers in the entire technology sector, creating a genuine competition for the machine-learning talent needed to build, test, and validate these systems, a competition the Department of Defense, bound by government pay scales and lengthy clearance processes, does not always win outright, which is itself part of the logic behind leaning on commercial contracts rather than building every capability in-house.
AI in defense
The phrase “AI in defense” invites a simpler story than the one actually unfolding. It is not one technology being adopted for one purpose. It is a frontier AI industry being wired into the world’s most sensitive computer networks under contested and still-evolving usage terms, a decade-old policy directive straining to govern targeting decisions that move at a pace its authors did not anticipate, a maintenance revolution quietly saving billions of dollars with almost none of the controversy attached to autonomous weapons, and a software layer, ASGARD, CJADC2, Sanctum, Lattice, increasingly doing more to determine battlefield outcomes than the hardware it happens to be running on.
What ties these threads together is not a shared technology so much as a shared constraint: every one of these efforts is racing against an adversary, a maintenance backlog, or a planning cycle that used to move at human speed and increasingly does not. The institutions built to govern, procure, and validate military technology, largely designed for an era when a new capability took a decade to field and a policy review could reasonably keep pace, are being asked to do the same job for a technology that changes meaningfully every few months. Whether that governance structure can be rebuilt fast enough to keep genuine human judgment meaningfully in place, across targeting, surveillance, and the eight corporate boardrooms now holding classified-network contracts alike, is arguably the more consequential question hanging over the next several years of military AI than any single new capability likely to be unveiled in that time.
