MUNICH — An HX-2 strike drone launched somewhere over Ukraine flies into heavy electronic warfare, keeps its target in view without a satellite link, and dives. The artificial intelligence on board was built by Helsing, a five-year-old defense firm now valued by investors at $18 billion. Its chief executive is a former biologist whose last company sold a physics engine to Grand Theft Auto.
That is the part of the AI arms race nobody advertises: the talent came from video games, the doctrine came from a commission chaired by the man whose company refused to build military AI, and the weapons procurement was shaped by a dogfight tournament where an eight-engineer contractor humiliated Lockheed Martin and a real F-16 pilot on the same afternoon.
Corporate filings, procurement records, government documents and on-the-record interviews map the whole pipeline — from a Oxford animation lab to a Pentagon task order, from a Google employee revolt to a CIA-adjacent venture fund, from an academic benchmark to a "Target Factory" that won a national prize. This is what the record shows.
Table of Contents
The tournament that was a shopping trip
In 2019, the Defense Advanced Research Projects Agency began running the AlphaDogfight Trials, a simulated dogfighting competition under its Air Combat Evolution program. On its face it was research: could machine-learning agents be trusted to fly fighter maneuvers? The stated purpose was larger — DARPA wanted to expand the developer base for air-combat autonomy.
Eight teams qualified. Among them was Lockheed Martin, the largest defense contractor in the Western world, alongside Aurora Flight Sciences (quietly a Boeing subsidiary), Perspecta Labs, the Georgia Tech Research Institute, SoarTech, EpiSci — and Heron Systems, a tiny software contractor from Alexandria, Virginia, that few in the audience could have picked out of a lineup.
In Aug. 2020, Heron's "Falco" agent defeated an experienced U.S. Air Force F-16 pilot 5–0 in simulated close-range gun duels. It also beat every other AI in the bracket, primes included. The tournament was not a tournament. It was a bake-off.
Eleven months later, Shield AI — the venture-backed firm of co-founders Brandon Tseng, a former Navy SEAL, and Ryan Tseng — acquired Heron Systems outright, folding its reinforcement-learning team into the Hivemind autonomy stack, which the company describes as an AI pilot that can fly without GPS or remote operators. In 2023, DARPA and the Air Force Test Pilot School flew an experimental F-16 — the X-62A VISTA — with AI agents at the controls in live tests that included AI-on-AI dogfights.
The acquisitions kept going. The Naval Air Systems Command awarded Shield AI a contract to put Hivemind on the Kratos BQM-177A — a subscale target drone, the aircraft used to simulate enemy threats in training exercises. The thing pilots used to shoot at now flies itself.
From benchmark to battlefield
flowchart TD
accTitle: How a DARPA research tournament became a procurement pipeline
accDescr: DARPA's ACE program sponsored the AlphaDogfight Trials, in which eight teams competed. Heron Systems won, was acquired by Shield AI, and its technology now appears in Shield AI's Hivemind deployments.
DARPA["DARPA"] -->|Sponsors| ACE["Air Combat Evolution (ACE)"]
ACE -->|Expands the developer base| Trials["AlphaDogfight Trials<br/>2019–2020 · simulated dogfights"]
Trials -->|Eight teams| Field["Lockheed Martin · Boeing / Aurora<br/>Perspecta Labs · Georgia Tech Research Institute<br/>SoarTech · EpiSci · Heron Systems"]
Heron["Heron Systems<br/>Falco AI"] -->|Defeated an F-16 pilot 5–0| Trials
Shield["Shield AI"] -->|Acquired in July 2021| Heron
Shield -->|Hivemind AI pilot| Kratos["Kratos BQM-177A target drone<br/>NAVAIR integration contract"]
Shield -.->|Company says its AI later flew| VISTA["Experimental F-16 flights<br/>2023 test program"]
Edges describe the competition, acquisition and contracts in the record. The dashed edge marks a later company-described milestone rather than a procurement decision.
From Grand Theft Auto to the kill chain
Torsten Reil trained as a biologist, studying neural networks and complex systems at Oxford. He founded NaturalMotion, an Oxford spin-off whose "euphoria" animation engine simulated how bodies fall, stumble and crash. Hollywood licensed it for The Lord of the Rings, Troy and Harry Potter. Rockstar Games licensed it for Grand Theft Auto IV and Red Dead Redemption — every spectacular car-crash ragdoll in those games runs on Reil's code. Zynga bought the company for $527 million in 2014.
That same year, Russia annexed Crimea. Reil has said the annexation is part of why he started again. In 2021, with Niklas Köhler and Gundbert Scherf, he co-founded Helsing.
Helsing says its software has been active in Ukraine since 2022 and that the core technology of its HX-2 — a mass-producible strike drone with onboard AI for finding targets through jamming — is deployed there. The company's public descriptions retain a human in the loop. In 2025, General Catalyst led a €450 million Series C with Accel, Lightspeed, Plural, Greenoaks and the individual investor Elad Gil, while legacy prime Saab came in as strategic investor and partner on Eurofighter electronic-warfare upgrades. By 2026, investors were backing a funding round reported to value Helsing at $18 billion, alongside a reported €1.46 billion German drone contract.
The career path — simulation, games, acquisition, defense unicorn — is not an anomaly. It is the industry's new labor market. Animated combat became real combat; the same engineers, the same optimization problems, the same black-box policy networks. Only the ragdolls changed.
The physics-engine pipeline
flowchart TD
accTitle: From a games and animation studio to an AI strike drone
accDescr: Torsten Reil's career path from Oxford spin-off NaturalMotion to Helsing and the HX-2 strike drone, with dates from corporate and press records.
Oxford["Oxford<br/>Biology and neural networks"] -->|Founded| NM["NaturalMotion<br/>Euphoria animation engine"]
NM -->|Physics for falling and crashing bodies| Films["The Lord of the Rings · Troy<br/>Grand Theft Auto IV · Red Dead Redemption"]
Zynga["Zynga"] -->|Acquired in 2014 for $527M| NM
Crimea["Russia annexes Crimea<br/>2014"] -.->|Reil cites it as motivation| Helsing["Helsing<br/>co-founded 2021 with Niklas Köhler<br/>and Gundbert Scherf"]
Helsing -->|Software claimed active since 2022| Ukraine["War in Ukraine"]
Helsing -->|Onboard AI for jamming resistance| HX2["HX-2 strike drone"]
GC["General Catalyst · Accel · Lightspeed<br/>Plural · Greenoaks · Elad Gil"] -->|€450M Series C| Helsing
Saab["Saab"] -->|Strategic investor and partner| Helsing
Helsing -.->|2026 round reported at an $18B valuation| Markets["Public and growth investors"]
Solid edges are corporate events and transactions in the record; dotted edges mark reported motivations and reported valuation figures.
The revolt that moved the work
In April 2017, the Pentagon quietly stood up Project Maven to run machine learning over tactical drone video for the Defeat-ISIS campaign. Google joined. In 2018, thousands of Google employees protested — some resigned — and Meredith Whittaker, then a researcher at the company, helped organize the internal campaign. Chief Executive Sundar Pichai published Google's AI principles, and Google Cloud declined to bid on a follow-on contract.
That is the story most people know. Here is the rest of the graph.
The contract did not die; it moved. Palantir Technologies now runs Maven Smart System, the production follow-on, and federal spending records list a $292.68 million task order with a five-year ordering period. The same firm powers Army Vantage, the service's data platform, and leads TITAN, the Army's first "AI-defined vehicle," for more than $178 million. The protest won a battle and emptied the field: the Pentagon learned to buy this work from companies whose employees do not get a vote.
Then the protesters' own industry crossed the line. OpenAI — founded in 2015 as a nonprofit to ensure artificial general intelligence benefits humanity — quietly deleted the "weapons development" and "military and warfare" bans from its usage policy in Jan. 2024. Several OpenAI employees reportedly learned about the change from a news article. In 2025, OpenAI announced a Chief Digital and Artificial Intelligence Office pilot worth up to $200 million. Anthropic, started by former OpenAI safety researchers, announced its own two-year Pentagon prototype agreement worth up to $200 million. Both arrive through Amazon Web Services and Microsoft cloud infrastructure built for classified and defense workloads.
Microsoft, meanwhile, states that it provides Israel's Ministry of Defense with software, professional services, Azure and Azure AI — while saying an earlier review found no evidence its systems were used to target or harm people and announcing a further review of allegations. Microsoft and Palantir separately offer integrated Azure OpenAI and Palantir AIP deployments for classified defense and intelligence networks. Google itself returned to the Pentagon fold in Dec. 2022, taking a share of the $9 billion Joint Warfighting Cloud Capability contract alongside Amazon, Microsoft and Oracle — four years after the Maven revolt, and around the time dozens of Google workers were fired for protesting its Project Nimbus contract with Israel.
By Dec. 2024 the loop had closed in public: Palantir, Anduril, SpaceX, OpenAI, Scale AI, Helsing and Saronic formalized a consortium to build a new defense-industrial stack. The company that refused to label drone footage now sits at a table with the company that runs the targeting software, next to the drones that feed it.
Whittaker resigned from Google in 2019. She is now president of the Signal Foundation — the encrypted messenger used, among others, by the people who would rather none of this were automated.
Before and after Project Maven
flowchart TD
accTitle: Where the military-AI work went after Google left Project Maven
accDescr: Google participated in Project Maven and exited in 2018 after employee opposition. The graph records the later Palantir follow-on program and the entry of OpenAI and Anthropic into Pentagon contracts.
DoD["U.S. Department of Defense"] -->|Established in 2017| Maven["Project Maven<br/>Defeat-ISIS drone video"]
Google["Google"] -->|Participated, then declined follow-on work| Maven
Revolt["Employee protests<br/>organized with Meredith Whittaker"] -.->|Adverse reaction| Google
Maven -->|Production follow-on| MSS["Maven Smart System<br/>$292.68M task order"]
Palantir["Palantir"] -->|Operates| MSS
Palantir -->|Also runs| AV["Army Vantage · TITAN<br/>$178M+ AI-defined vehicle"]
OpenAI["OpenAI"] -->|Deleted its military-use ban, Jan. 2024| Policy["Usage-policy change"]
Policy -->|Followed by| Pilot["$200M CDAO ceiling pilot"]
Anthropic["Anthropic"] -->|Two-year prototype agreement| DoD
Google -->|2022 JWCC cloud award, $9B vehicle| JWCC["Joint Warfighting Cloud Capability<br/>with Amazon · Microsoft · Oracle"]
Cons["December 2024 consortium<br/>Palantir · Anduril · SpaceX · OpenAI<br/>Scale AI · Helsing · Saronic"] -.->|Formalized partnership| AV
The dashed edges mark reactions and partnerships rather than procurement decisions. The policy node records OpenAI's own usage-policy text change.
The man who wrote the shopping list
Eric Schmidt ran Google through the years its unofficial motto was "don't be evil." He then made a second career out of arguing that America was not taking the AI arms race seriously enough.
In 2018, Congress created the National Security Commission on Artificial Intelligence and installed Schmidt as chair. His vice chair was Robert O. Work, the former deputy secretary of defense who had bet the Pentagon's modernization on a "third offset" of autonomy and AI after Russia's 2014 moves in Ukraine. Its executive director was Ylli Bajraktari. The commission's 2021 final report ran 750 pages and told Congress to spend at wartime speed — on data, talent, autonomy and adoption. Congress largely listened: the commission's organizational blueprints echo through the offices that now write the checks.
Schmidt then founded the Special Competitive Studies Project to keep the pressure on, publishing a 2024 report with the Royal United Services Institute arguing that human-machine teaming should sit at the center of U.S. and U.K. deterrence. And in Jan. 2024, Forbes revealed what the chair of the commission on responsible AI competition had done next: Schmidt had quietly founded White Stork, a stealth startup building AI-guided "kamikaze" drones — with a design target of roughly $400 per drone, incorporated in Estonia, and reportedly test-flown in a wealthy Silicon Valley suburb. Reports in 2026 describe the drones, developed with a cluster of small engineering firms, achieving autonomous hit rates above 70 percent in Ukraine — a figure that remains a company-linked claim rather than an audited one.
Read the sequence again: the man whose company refused to help the Pentagon sort drone footage chaired the commission that told Congress to industrialize military AI, and then went into the kamikaze-drone business himself. The distance between "we don't do weapons" and a $400 loitering munition is one career.
The Schmidt pipeline
flowchart TD
accTitle: Eric Schmidt's path from tech executive to commission chair to drone manufacturer
accDescr: Roles and events in the record, from Google through NSCAI and SCSP to the reported White Stork drone venture.
Schmidt["Eric Schmidt"] -->|Chief executive| Google["Google<br/>Declined follow-on Maven work in 2018"]
Schmidt -->|Chair, 2018–2021| NSCAI["National Security Commission<br/>on Artificial Intelligence"]
Work["Robert O. Work<br/>Third-offset advocate"] -->|Vice chair| NSCAI
Bajraktari["Ylli Bajraktari"] -->|Executive director| NSCAI
NSCAI -->|2021 final report| Congress["Congress<br/>Funding and organizational recommendations"]
Schmidt -->|Founded| SCSP["Special Competitive Studies Project"]
SCSP -->|2024 report with RUSI| HMT["Human-machine teaming<br/>as deterrence doctrine"]
Schmidt -.->|Reported in 2024 as founder| WS["White Stork<br/>AI-guided kamikaze drones<br/>~$400 design target"]
WS -.->|Reported testing and deployment| Ukraine["Ukraine"]
Solid edges are on-the-record roles and publications. Dotted edges mark the reported stealth venture and its reported use, as covered by Forbes and trade press.
Inside the Target Factory
Nowhere is the accountability gap wider than in the Israel-Gaza war, and the naming conventions alone should give a reader pause.
Israel's Ministry of Defense publicly describes the Target Factory — winner of the Israel Defense Prize in 2024 — as a system combining cross-branch information sharing, "advanced algorithms and artificial intelligence" to accelerate target planning and strike preparation. The IDF confirms it uses Habsora (the Gospel) for data collection and processing and Lavender as a database supporting analysts, and it rejects any description of autonomous target selection.
Investigative reporting draws a darker diagram. Yuval Abraham, reporting for +972 Magazine and Local Call, described Lavender as a system that ranked suspected militants for targeting with minimal human review, and Where's Daddy? as a tracking tool that alerted when a target reached a family home. Human Rights Watch has warned that digital tools assigning suspected-affiliation ratings could reproduce bias and raise civilian-targeting risks. The most consequential claims remain contested — reported and alleged, not adjudicated.
Two facts sit above the dispute. First, the industrial metaphor: a military publicly collects a national prize for a "factory" that manufactures targets at speed, while journalists investigate whether the assembly line eroded the meaning of human judgment. Second, the supply chain: Microsoft states it provides Israel's Ministry of Defense with Azure and Azure AI services under its cloud and support contracts. Whatever the review of targeting allegations concludes, the infrastructure question is separate and already answered — commercial cloud and frontier models are now standard components of the targeting stack.
Two stories about the same software
flowchart TD
accTitle: Competing public accounts of Israeli AI-assisted targeting systems
accDescr: The official IDF and Israeli Defense Ministry account and the investigative account reported by +972 Magazine and Human Rights Watch are shown as separate evidence paths with their confidence levels.
TF["IDF Target Factory<br/>Israel Defense Prize 2024"] -->|Official description: algorithms and AI<br/>for target planning and strike preparation| Plans["Target planning and strike preparation"]
IDF["Israel Defense Forces"] -->|Says: intelligence processing<br/>and analyst support| Habsora["Habsora / the Gospel"]
IDF -->|Says: a database for analysts| Lavender["Lavender"]
Journo["+972 Magazine · Yuval Abraham<br/>Human Rights Watch"] -.->|Reported: rank-ordered suspected militants<br/>with limited human review| Lavender
Journo -.->|Alleged: tracking tool tied<br/>to strike timing| WD["Where's Daddy?"]
WD -.->|Alleged association with| Gaza["Israel-Gaza war and<br/>civilian-harm reporting"]
Lavender -.->|Contested operational role| Gaza
Cloud["Microsoft Azure and Azure AI<br/>software and services to Israel's MOD"] -->|Company statement: earlier review found<br/>no evidence of use in targeting or harm| IDF
Cloud -.->|Further review announced in 2025| Dispute["Unresolved allegations"]
Solid edges are official statements and corporate disclosures. Dotted edges mark reported or alleged claims that are contested and unresolved.
The kamikaze economy and the attritable future
In Aug. 2023, Deputy Secretary of Defense Kathleen Hicks announced Replicator: field multiple thousands of "all-domain attritable autonomous systems" within 18 to 24 months, on roughly $500 million of initial funding. Read the adjective again. Attritable is procurement language for disposable — machines designed cheap enough to lose. A Sept. 2024 memorandum added Replicator 2, aimed at production capacity and at removing integration, policy and force-structure barriers. Scale, not sophistication, is the doctrine.
Ukraine is where the doctrine gets its feedback loop. The Ministry of Defence's A1 center turns combat data into forecasting, autonomy and command-and-control software; the DELTA situational-awareness system and the UA DroneID initiative — signed with the small security firm Cossack Labs among others — aim to identify drones in the air to stop soldiers shooting down their own machines. The U.K.-led Drone Coalition plans up to one million first-person-view drones, explicitly including AI-upgraded models, with participating states pledging tens of millions of euros to a common fund. The U.K. Ministry of Defence's 2024 drone strategy draws explicit lessons from that battlefield and commits at least £4.5 billion over a decade.
Then the boundaries blur. In 2024, the AUKUS partners — the United States, United Kingdom and Australia — ran real-time trials in which AI-enabled uncrewed aircraft from three countries exchanged data and supported human-operated target-location and strike workflows. MBDA, the European missile group, is testing Orchestrike, in which cruise missiles collaborate with each other and their pilot in flight — the company's phrasing, and its assurance that a human operator remains in control. Its Sky Warden counter-drone system uses AI to classify threats and assign which weapon should shoot.
Games taught the industry the phrase "human in the loop," and games taught it the trick of making the loop feel like a controller in your hands. Missiles that talk to each other, drones that hunt through jamming, and a procurement target of thousands of disposable machines is what "human-machine teaming" looks like in hardware.
The mirror in Beijing
While Washington built its AI acquisition stack, Beijing built a mirror. In 2017 — the same year Project Maven began — China's State Council issued its New Generation Artificial Intelligence Development Plan, and Xi Jinping personally chaired the new Central Commission for Military-Civil Fusion Development, charged with wiring civilian technology into defense modernization. U.S. Defense Department reporting describes the People's Liberation Army pursuing "intelligentized warfare": AI-enabled command and control, sensor fusion, decision aids and autonomous systems.
The scale is now measurable. A 2025 study by Georgetown's Center for Security and Emerging Technology analyzed 2,857 AI-related PLA award notices from 2023–2024, spread across 1,560 winning organizations — a procurement base mixing state-owned champions like CETC, CASC and NORINCO with universities, research institutes and civilian vendors. SenseTime, Megvii and iFlytek sit in this dataset as research leads, not proven contract relationships — the military links of some civilian AI firms remain a follow-up question, and this graph labels them that way.
India's defense research organization, DRDO, runs its own path: the Centre for Artificial Intelligence and Robotics working on command-and-control and autonomous systems, and centers of excellence at the Indian Institute of Science and IIT Jodhpur for AI-enabled information warfare, war gaming and UAV swarms.
The uncomfortable symmetry is that both systems arrived at the same place — dual-use AI, civilian labs, procurement as industrial policy — from opposite directions. In America, private capital and defense demand pulled the labs in. In China, state direction pushed the labs out. The gap between the two systems is no longer technical. It is a factory problem.
The rules written after the fact
The normative architecture arrived as a polite trailing edge. A 2015 open letter hosted by the Future of Life Institute — signed by Stephen Hawking, Elon Musk and DeepMind's Demis Hassabis, among thousands — warned that a military AI race would make autonomous weapons cheap and ubiquitous. A 2018 pledge declared that decisions to take human life should never be delegated to machines. The International Committee of the Red Cross has pressed the CCW group of governmental experts for prohibitions on systems designed to target people and strict human-control limits on the rest. The Campaign to Stop Killer Robots, co-founded by Human Rights Watch, now spans more than 250 organizations in over 70 countries demanding a binding treaty.
States, predictably, produced declarations. The Netherlands and South Korea convened the first global summit on responsible military AI at The Hague in 2023 (REAIM), where the U.S. Department of State launched the Political Declaration on Responsible Military Use of AI and Autonomy — non-binding commitments to auditable, tested, deactivatable systems — with 45 endorsing states by that November. Seoul hosted REAIM 2024. And in Dec. 2024 the U.N. General Assembly adopted Resolution 79/239, its first resolution on AI in the military domain, by 159–2–5.
None of these instruments is a treaty. A $292.68 million task order, a €450 million Series C and a $18 billion valuation are contracts. The money is binding; the rules are not. Between the two sit the acquisition lawyers, and they have opinions about which document matters.
What the graph shows
Strip the narrative away and the structure is stark. Of 161 entities and 131 relationships, the highest-severity edges cluster in one place: the contested targeting systems in Gaza, where official and investigative accounts of the same software diverge. Everything else — the strategies, the summits, the venture rounds — is severity 2 and below: institutional, financial, procedural.
The bridge nodes are just as legible. CDAO links strategy to procurement; Palantir links Maven, Army Vantage, TITAN, Microsoft and classified networks; Anduril links venture capital to the edge-data mesh and the scale-production agenda. Eric Schmidt links frontier technology to policy to manufacturing. And two investors — General Catalyst and the individual Elad Gil — appear on both sides of the Atlantic, in Anduril's Series D and Helsing's Series C alike. One cap table spans both wars.
Trae Stephens — Anduril co-founder and Founders Fund partner — is in this graph too, and in this publication's earlier investigation of American police surveillance technology. The same few nodes finance the street-pole camera and the strike drone. Readers of The Glass Dragnet have already met half the cast of the kill chain; this is simply the other side of their balance sheet.
Four questions now deserve investigators' attention. First, what human-control architecture actually governs Lavender-class targeting pipelines, and who audits the difference between a "database" and a ranking system? Second, what data-rights, model-update and incident-reporting terms govern the frontier labs now inside Pentagon and Ministry of Defense clouds? Third, how will "attritable" autonomy be tested for discrimination, fratricide and override when communications are denied? Fourth, when a strike decision passes through a game developer's policy network, a venture fund's subsidiary and a cloud giant's service agreement — where does the accountability legally live?
The 2015 open letter's fear was a military AI arms race. It has arrived, and it looks nothing like the movies: no chrome terminators, just contracts, procurement portals, and a former games programmer explaining that a human presses the button. The machine that picks the target was built by the best-staffed engineering culture in history — hired away from the companies that taught the world to play.
Sources and further reading
- DoD CIO, Project Maven establishment memorandum (April 2017).
- USAspending, DoD Palantir USG IDIQ, Maven Smart System ($292.68M task order).
- Pentagon announces Joint Warfighting Cloud Capability awards to Amazon, Google, Microsoft and Oracle (Dec. 2022).
- Shield AI, Shield AI Acquires Heron Systems (July 2021); Over Defense, AlphaDogfight Trial Winner Heron Systems Acquired By Shield AI.
- DARPA, Air Combat Evolution (ACE) and AlphaDogfight Trials program materials; Defense Advancement coverage of the NAVAIR/Kratos BQM-177A integration.
- NaturalMotion, corporate history and euphoria engine licensing; NaturalMotion; Zynga acquisition ($527M, 2014).
- Helsing, company and HX-2 disclosures; Ukraine War Analytics, Helsing: Europe's AI Defense Unicorn and the Ukraine War; BattlePolicy, Helsing Raises $1.8 Billion at $18 Billion Valuation (2026).
- Sarah Emerson, Eric Schmidt's Secret Military Project Revealed: Attack Drones, Forbes (Jan. 2024); testing in a wealthy Silicon Valley suburb (June 2024).
- NSCAI, Final Report (2021); SCSP–RUSI, Leveraging Human-Machine Teaming (2024).
- OpenAI quietly removes ban on military use of its AI tools, CNBC (Jan. 2024); WIRED on the Pentagon's testing of OpenAI models.
- Israeli Ministry of Defence, Israel Defense Prize 2024, including the 'Target Factory' project.
- IDF statement on its data technologies; Yuval Abraham, +972 Magazine investigation of Lavender; Human Rights Watch, Q&A on the Israeli military's use of digital tools.
- DoD, Deputy Secretary Kathleen Hicks, "The Urgency to Innovate" (Replicator); CDAO/DIU, AI Rapid Capabilities Cell announcement.
- U.K. Ministry of Defence, Defence Artificial Intelligence Strategy; AUKUS real-time AI and autonomy trials (Dstl, 2024).
- Ukraine Ministry of Defence, Defense AI Center A1 and DELTA/UA DroneID materials.
- CSET, Pulling Back the Curtain on China's Military-Civil Fusion (2025).
- U.S. State Department, Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy; UNGA Resolution 79/239 (Dec. 2024); ICRC autonomous-weapons recommendations.
- Companion investigation: The Glass Dragnet.