Ukraine Opens Its Drone-Footage AI Archive to Allies Free, and the Price Is a License Back to Kyiv

With Avengers Labs, its annotated archive of battlefield imagery, Kyiv is trying a different kind of trade. Allied defense companies can train artificial-intelligence models on the archive for free. The price is the software: whatever a company builds there has to be licensed back to the Ukrainian military.

The arrangement was described in a New York Times report on November 12, relayed by DroneXL. According to that account, access costs allied firms nothing up front, and up to 12 British companies are first in line. Ukraine is also in talks with other governments. The Times also reports something the official announcements leave out: the licence the companies sign contains no human-rights clause.

What is actually in the archive

Ukraine's Ministry of Defence describes the dataset as more than 5 million annotated frames, mostly gathered from unmanned systems through DELTA, the military's situational-awareness system. The ministry says data of comparable volume and quality is virtually impossible for a private company, or a country without experience of modern full-scale war, to collect.

Ukrainian developers already use it. The ministry says more than 30 Ukrainian companies train on the archive and that three products built this way are now in combat use. One of them is a detection system that processes more than 100,000 drone video streams a month.

The archive's reputation rests on its performance numbers, and those need care. In an August 10 release cited by Janes, the ministry said the platform had trained a target-detection system that handles more than 100,000 UAV video streams monthly and detects 70% of all targets in real time. The Times report adds that an executive, speaking anonymously, put the false-positive rate below 5 percent. Neither figure has been independently verified. Both come from the ministry or from an unnamed industry source. Neither source defines a "target," the conditions behind the 70% figure, or how a false positive is counted, and those details determine whether the numbers are impressive or just reasonable.

How the sandbox works

Ukraine is not handing over the footage itself. According to the ministry, participants cannot download the dataset and must train inside a secure perimeter. The Times account goes further: companies cannot even view the footage. The archive is encrypted and hosted with a cloud service outside Ukraine.

This is a "data stays home, models travel" approach. A foreign engineer can run training jobs against the corpus and take away weights, but has no access to the raw video. That raw video could reveal Ukrainian positions, tactics, unit identities or the faces of soldiers and civilians. For Kyiv, the design keeps operationally sensitive material under its control while still getting allied engineering work done on it. For participating firms, it means building models on data they cannot inspect, which makes it harder to audit labeling errors, bias in the annotations, or gaps in what the frames cover.

Access is also not automatic. The ministry says the first competition under the British arrangement seeks AI models that can operate when contact with operators or GNSS satellite navigation is limited. The UK Ministry of Defence's RAID Task Force will select participants within weeks, and access opens only after Ukrainian export-control procedures. The ministry lists AI models for swarms of unmanned systems as the priority.

How Britain got first access

The British deal rests on a government-level agreement. Janes reports that a UK-Ukraine AI partnership agreement was signed on August 24. According to Defense News, that deal was signed in Kyiv by UK Prime Minister Andy Burnham and President Volodymyr Zelensky. The Ukrainian ministry says the Avengers Labs access rests on that agreement together with a joint declaration of intent.

On August 27, a defence source told Janes that the UK was considering using Avengers Labs data "for decision support and target identification functions." Janes also reported two early projects involving UK firms Sintela, Mind Foundry and Skyral: one using fibre-optic cables as AI sensors, the other developing low-power AI chips for unmanned systems. By late September, Defense News reported that the British MoD had struck a partnership with Avengers AI Labs. London officials described it as the first time Kyiv had opened its wartime data to an outside government.

The NYT report, as relayed by DroneXL, says the British arrangement is still the only agreement in place, although Kyiv is negotiating with other governments.

Why It Matters

The deal has three implications.

It is a new export model. Ukraine is turning combat experience into a resource it can trade. Instead of selling the data or giving it away, it lends access and takes the derived software in return. Allied firms get a training corpus no peacetime program could reproduce, and Ukraine gets models it did not pay to develop, built by engineers it does not employ. If other governments sign on, Ukraine could end up with a small portfolio of foreign-built AI tools, each licensed back to its own forces. That gives Kyiv more leverage in arms diplomacy and adds a steady stream of software to its defense base.

It points at autonomy, not just analytics. The first stated requirement, models that keep working when operator links or GNSS are degraded, describes drones that must perceive and decide with less human supervision. A swarm priority points the same direction. Janes' source talking about "target identification" makes the link explicit. These are not tools for sorting archived imagery. They are perception layers for weapons.

The governance has gaps. If the Times is right that the licence carries no human-rights clause, the main contract governing these models says nothing about how they may be used. According to DroneXL, Ukrainian officials say each receiving government must make its military follow the Geneva Conventions and other treaties, so the question falls to national export controls, the UK selection process, and international humanitarian law. None of those was clearly designed for a case where a company trains a model on footage it cannot see, inside a foreign government's sandbox, and the model then goes to that government's military. Accountability is also hard to assign. The developer cannot audit the training data, the data owner did not write the code, and the operator in the field may not know where either came from. The headline accuracy figures, which are unverified, show why this matters: a 5 percent false-positive rate means something very different for a dashboard than for a seeker that commits a munition.

None of this makes the arrangement unusual for wartime procurement, which rarely waits for ethics frameworks to catch up. It does mean that as Kyiv brings in more partner governments, the terms of the licence will matter as much as the size of the dataset. Those terms are who owns the models, who can modify them, and whether they carry any limits on use.