Ukraine Opens 5 Million Real Combat Photos to the UK First

Why These 5 Million Photos Matter

When Western defense contractors train AI to read the battlefield, most of what they use is synthetic data built through simulation, because footage from real combat has simply been unobtainable. That's now changing. Ukraine has opened up 5 million real combat photos it has collected via drones and sensors since 2022, and the UK is the first country to get access. UK Prime Minister Andy Burnham and Ukrainian President Volodymyr Zelensky signed an AI partnership in Kyiv, opening Ukraine's battlefield data platform, Avengers AI Labs, to British researchers and companies. According to the Financial Times, this is the first time any foreign government has been given access to this data, and the deal falls under the "100-year partnership" the two countries have signed.

The project dates back to 2022, when Ukrainian drone units began training neural networks on footage they'd captured, trying to automatically spot Russian troops and vehicles. The goal was to speed up the military's decision cycle — observe, orient, decide, act. Until now, this data had only been shared with domestic companies, which is one reason Ukrainian drone makers have outperformed foreign competitors trained on synthetic data when it comes to real-world recognition. Ukraine signaled back in March that it would start sharing this data with allies, and this agreement with the UK is the first instance of that promise being kept.

Inside the Data Room Only — Rules Built With Palantir

According to Ukraine's Ministry of Defense, the core of this deal is a set of roughly 5 million labeled images, most of them drawn from DELTA, the digital battle-management system that stitches together real-time footage from drones, satellites, and sensors. Approved companies can only train and test computer vision models inside a secure data room built jointly with Palantir — the finished AI models remain Ukraine's property. Rather than exporting the data, the arrangement grants access rights to where the data already sits.

One detection system already built on this data offers a glimpse of what it can do. According to the Ministry of Defense, it identifies 70% of enemy equipment appearing in video feeds, and it takes just 2.2 seconds to process each object. The system reportedly handles more than 100,000 drone video streams every month.

Metric	Figure
Labeled image dataset	~5 million
Monthly drone footage processed	100,000+
Enemy equipment identification rate	70%
Processing time per object	2.2 seconds
Interceptor drone autonomy level	~95%

What Three UK Startups Are Building

According to a UK government announcement, three startups — Sintela in Bristol, Mind Foundry in Oxford, and Skyral in London — already have pilot projects underway. One turns buried fiber-optic cables into AI-powered sensors that can protect military bases and, eventually, airports and railways. Another is developing low-power AI chips for drones and autonomous systems. British forces are especially interested in acoustic sensor data for detecting incoming Russian drones, which, properly trained, reportedly delivers far higher accuracy than radar. The announcement came just after Burnham revealed he would allow classified information about British-made components in MBDA's SCALP cruise missiles to be shared with Ukrainian assembly lines.

How Far Drone Autonomy Has Come

What this data is actually building becomes clear when you look at the three stages of drone autonomy: autonomous navigation, in which a drone finds its own route without GPS; "last-mile" tracking, in which a human designates a target and AI handles the final approach; and autonomous target selection, in which the machine itself chooses what to strike.

The first two stages have already been standard on Ukrainian battlefields for some time. In 2024, a Ukrainian FPV drone that lost its radio link reportedly struck a pre-designated Russian tank on its own. According to the government announcement, Ukrainian interceptor drones now operate with roughly 95% autonomy, and software from Swarmer has coordinated drone swarms across more than 100 combat missions. Swarmer's Misha Nestor told the Financial Times that "high-quality labeled battlefield data is the single biggest bottleneck to developing reliable autonomous systems."

What Happened in Zaporizhzhia

The third stage — machines choosing their own targets — has drawn fresh scrutiny thanks to a New York Times report. Last July, a Russian Molniya drone killed three civilians in Zaporizhzhia, including 19-year-old university student Tetiana Bubinets. The operator had originally programmed the drone to strike a gas station, but investigators found that the software instead chose to hit what appeared to be a propane tank nearby, on its own.

Recovered wreckage contained a commercial NVIDIA Jetson Orin mini-computer, which appears to have played a role in target selection. Kateryna Bondar of CSIS called the incident the first documented case of a Russian drone equipped with autonomous targeting AI causing civilian deaths. Investigators cautioned, though, that the presence of an NVIDIA chip alone doesn't prove autonomous decision-making — that conclusion depends on the lost radio link to the operator lining up with the code and training data found on the computer.

Ukraine isn't exempt from this trend either. Mykhailo Fedorov, recently removed as defense minister, told the New York Times that Ukraine had tested fully autonomous AI systems for several months in occupied Crimea, striking fuel depots and military equipment without civilian casualties. From the Saker Scout autonomous attack drone introduced in 2023 to today's drone swarms, Ukraine has steadily expanded its use of autonomous weapons. As recently as April 2026, assessments held that fully autonomous weapons didn't exist on Ukrainian battlefields — but the tests that have since come to light are forcing a reassessment of that conclusion.

Editor's Take

Reading this deal as just another data-sharing agreement misses half the story. The labeled data Ukraine has accumulated in real combat over more than three years is something synthetic data could never replicate, and the moment it went from domestic-only to shared-with-allies, one of the biggest bottlenecks in global defense AI development quietly disappeared. Western drone makers have long trained their models on fake battlefield footage from simulations or game engines, and there's been persistent criticism that this hurts real-world recognition performance. This access shortens that gap considerably.

The generational leap is easier to see in comparison. Back in 2022, when Ukraine first ran neural networks on drone footage, the goal was simply "automatically spot troops and vehicles." Now there's a deployed system that processes objects in 2.2 seconds with a 70% identification rate, and on top of that, stage-three autonomy — machines picking targets without human input — is being confirmed on real battlefields. In four years, the center of gravity in autonomy has shifted entirely from detection to decision-making.

Practically speaking, this is a signal that any company looking to get into defense AI can't afford to ignore. Training exclusively inside a secure data room without exporting the underlying data, while letting the data-owning country retain ownership of the finished AI, looks likely to become the standard contract structure for future defense AI partnerships between nations. Conversely, if a company has built its computer vision models purely on synthetic data, this is the moment to seriously reckon with the widening gap between it and competitors who now have access to real combat data.

Expect more controversies like the Zaporizhzhia incident over the coming months as autonomous targeting spreads. International norms for autonomous weapons still haven't caught up with the pace of real-world deployment — and in the meantime, both the data and the systems built on it are already moving to the next stage.