Ukraine’s battlefield AI is exposing a dangerous surveillance gap that could reach Australian citizens through AUKUS, writes Professor Vince Hooper.
WHEN A GOVERNMENT describes a foreign military archive as a ‘goldmine’, it is worth paying attention to the metaphor. Goldmines produce wealth for the owner of the processing plant, not for the earth that yields the ore.
On Monday, UK Prime Minister Andy Burnham stood in Kyiv on Ukraine’s Independence Day and signed a deal that makes Britain the first international partner admitted to Ukraine’s Avengers AI Labs.
The platform sits on top of Delta, a battlefield management system where a neural network trained on 5 million annotated images decides, roughly 70 per cent of the time, which of the objects crawling across more than 100,000 monthly drone feeds is a Russian tank and which is a tractor.
Built from four years of daylight cameras and infrared sensors stretched along the longest land war in Europe since 1945, Avengers is, by any measure, the most comprehensive real-world military AI training dataset ever assembled. Burnham called it a partnership. His government’s choice of metaphor told a different story.
Three British startups, Bristol’s Sintela, Oxford’s Mind Foundry and London’s Skyral are already running pilot projects. The headline application is domestic: buried fibre-optic cables turned into AI-enabled sensors to protect UK military sites from protesters and hostile state actors. Further use cases include airports, railways, prisons and energy infrastructure. In return, Ukraine gains access to British university research and commercial engineering, shortening the cycle between identifying a battlefield problem and deploying a technological fix.
On its face, the exchange looks elegant. Ukraine holds what amounts to a monopoly on annotated, real-world battlefield AI training data. No NATO exercise, no synthetic dataset comes close. Models trained outside Ukraine have mostly worked from openly published material, and the gap between open-source imagery and combat-annotated sensor feeds is roughly the gap between a textbook cadaver and a patient on the table.
That should worry Australians.
AUKUS Pillar II, the technology cooperation arm of the trilateral pact, designated AI-enabled payloads for uncrewed undersea vehicles as its first signature project in May. Defence Minister Richard Marles has flagged between 12 and 15 billion Australian dollars over the next decade for autonomous capabilities. Australian soldiers tested drone swarming tactics alongside British and American counterparts on Salisbury Plain this June. In February, the Maritime Big Play exercise on Australia’s east coast demonstrated live strikes by autonomous one-way attack munitions.
The appetite for battlefield AI data in Canberra is not hypothetical; it is budgeted.
Yet Australia was not in the room on Monday. The UK-Ukraine deal is bilateral and Canberra’s access to Avengers Labs, if it comes at all, will likely travel through the Five Eyes architecture or via a separate memorandum.
But Canberra may not need its own deal. Under AUKUS Pillar II, the three partners share AI research, autonomous systems technology and the training data that feeds them. If British engineers train models on the Avengers dataset and those models are integrated into jointly developed AUKUS platforms, Ukrainian battlefield data enters Australian defence systems through a trilateral framework that was never designed to handle third-party wartime data.
Nobody in Canberra has to approve the transfer because it arrives as capability, not intelligence. A week after the Avengers deal was signed, Marles flew to Washington and London to discuss AI and AUKUS with his American and British counterparts, talking publicly about battlefield data and AI-operated platforms. Back in Canberra, the University of NSW (UNSW) is already building a purpose-built defence and cyber security precinct on Constitution Avenue, within the national defence and security corridor, with its first cyber security cohort enrolled and two new buildings due to open in 2028.
The governance gap is not prospective; it is architectural, and the negotiation is already underway.
Every surveillance technology in modern history has followed the same migration path: battlefield to border to city centre. Radar became air traffic control. Signals intercepts became metadata collection. Drone cameras became police helicopters. The journey from Kandahar to Kansas took roughly a decade; the journey from Donetsk to Dartford is being compressed into a single procurement cycle.
The Guardian‘s reporting on the Burnham-Zelenskyy deal made the landing point explicit: AI models trained to distinguish a Russian T-72 from a Ukrainian combine harvester will now be calibrated to detect protesters at British defence installations. Sintela’s distributed acoustic sensing technology, already deployed along the U.S.-Mexico border under a $200 million contract with Customs and Border Protection, can identify footsteps, fence-cutting and digging in real time.
When the same fibre-optic infrastructure is laid beneath a British military base, “detection” of “protest activity” ceases to be a metaphor for security and becomes a literal description of surveillance.
The danger is not that the technology fails. It is that training data defines what counts as a threat and a dataset assembled in a war zone has one overwhelming prior: everything that moves is hostile until proven otherwise. A model trained to flag a T-72’s thermal signature against Ukrainian steppe will behave unpredictably when pointed at a delivery van near a British naval base.
Steven Murdoch at University College London (UCL) noted that the underlying engineering, detecting movement through fibre-optic cable, is decades old. He was not convinced Ukraine’s contribution changes the picture. But what Ukraine adds is scale: four years of annotated feeds from a theatre where the targets are real. But the targets on a British perimeter fence are not real in the same sense. Protesters are not drones. A dataset that cannot distinguish between the two is not a capability; it is a liability dressed in khaki.
Australia has been here before, or close to it. Pine Gap already operates as a signals intelligence hub under the UKUSA Agreement. The Australian Federal Police (AFP) trialled Clearview AI‘s facial recognition database, initially denied it, then were exposed when a stolen customer list surfaced. The Privacy Commissioner found that Clearview had breached Australian law.
The AFP subsequently met with the company again. The pattern – adopt first, govern later, deny in between – is precisely the pattern that a deal like the Avengers partnership invites at scale. The 2026 National Defence Strategy offers little guidance on balancing rapid capability development with escalation management and if the data is already flowing through AUKUS channels before anyone drafts an oversight framework, the framework arrives as decoration rather than architecture.
The deal also reveals a structural asymmetry that the “goldmine” metaphor, to its credit, makes legible. Britain can train and retrain models on the data indefinitely, developing capability at a time and specification of its choosing. Ukraine’s side of the bargain is more fragile: the promise that Western research and development will accelerate the drone-detection systems it desperately needs on the front line.
That promise holds only as long as the war holds Western attention. The asymmetry favours the partner that owns the processing plant, not the partner that produced the ore. If the war ends or Western interest wanes, Ukraine is left having opened its most sensitive military archive to a country that no longer needs to reciprocate. In the extractive industries, they call that a depleted mine.
Australia, if it enters this arrangement, whether by front door or through the AUKUS pipeline, should do so understanding what it is acquiring. Not just capability, but a wartime surveillance grammar: thermal signatures calibrated for the Donbas steppe, behavioural patterns drawn from a front line where civilians have long since fled, threat classifications forged in a context where ambiguity is lethal.
None of that translates cleanly to a railway station in Sydney or a naval base in Western Australia, where the person walking toward the perimeter may be an intelligence operative, a lost tourist, or a citizen exercising a constitutional right. Before the memorandum is signed, someone should ask whether the fibre-optic cable beneath Pine Gap will one day listen for the footsteps of an Australian citizen walking to a protest and whether the model that decides what those footsteps mean was trained in a place where the only right answer was “threat”.
The Avengers, in the original comic, were assembled to fight a battle too large for any single hero. Ukraine’s Avengers Labs was assembled to fight a battle too large for any single army. The question now is what happens when the data from that battle is reassembled in a country where the enemy is not an invading armoured column but a citizen with a placard.
Nobody in Kyiv trained the model for that.
Professor Vince Hooper is a proud Australian-British citizen and professor of finance and discipline head at SP Jain School of Global Management with campuses in London, Dubai, Mumbai, Singapore and Sydney.
Support independent journalism Subscribe to IA.
Related Articles