As artificial intelligence (AI) increasingly begins to work its way into military hardware, developing effective countermeasures is becoming increasingly urgent. With Ukraine’s increasing integration of ‘machine vision’ targeting capabilities into drones, it is only a matter of time before Russia catches up. In preparation, Ukraine has already begun examining counter-AI techniques.
At the Association of Old Crows (AOC) Europe 2026 conference, held from 19-21 May in Helsinki, Warsight sat down for an interview with Iaroslav Kalinin, CEO of Infozahyst, a Ukrainian company specialising in electronic warfare (EW), to discuss the latest developments on the battlefields of Ukraine.
The discussion centred on the bleeding edge of the rapidly-evolving drone war, including the rise of artificial intelligence (AI) in targeting and the developing field of counter-AI, as well as recent technical and tactical adaptations from Russia and Ukraine.
Over the past several years, AI has been rapidly making its way into military usage, from command and control (C2) systems, battle management systems (BMSs), sensor-to-shooter systems, navigation systems, to tools for jamming, electronic intelligence (ELINT), signals intelligence (SIGINT), as well as analysis and planning software.
In Ukraine’s drone industry, AI has begun to work its way into targeting, in the form of ‘machine vision’, in which on-board AI uses the drone’s camera(s) to see the target and taking over in the terminal phase to ensure the munition strikes its target accurately. AI targeting has already been employed by both Russia and Ukraine since at least 2025, largely for engaging ground targets, but efforts to integrate the technology into interceptors are also well underway.
According to Kalinin, the primary use of AI for targeting was “not resisting jamming”, but rather about “last-mile efficiency”. Kalinin explained in reference to the problem: “It’s a problem of mastery of [the] operator. So human reaction is [the] problem, especially if you’re utilising mesh network or satellite connectivity. Because of latency, you have to have huge mastery to actually efficiently operate this type of drone. If [the] target is manoeuvrable it becomes [difficult].”
At short ranges, as seen with typical first-person view (FPV) drones operating over radio links, latency is not particularly problematic, but the problem changes somewhat with distance, especially when having to bridge two different network connections, such as a satellite communications (SATCOM) link to a ‘mothership’ drone bridged with a radio link to an FPV drone launched from the mothership. Strike packages of this type have been used to conduct long-range precision strikes against key targets such as air defence system components, for instance.
When conducting accurate strikes at very long ranges, Kalinin explained, latency can become a problem for “any type of connectivity, it doesn’t matter if we talk about mesh network or we talk about the satellite network, or we talk about the MANET [Mobile Ad-hoc Network] network where satellite and mesh is combined.”
However, Kalinin noted that there are exceptions, such as LTE (4/5G cellular) technology: “LTE is quite good from the latency perspective. You could achieve up to 100 milliseconds, which is good.” However, control via LTE is not always possible, as use of such technology tends to rely on the presence of cellular infrastructure, whose operation can be blocked by governments, as was the case with Russia, where mobile connectivity has experienced periodic outages for much of early 2026. As such, bridging SATCOM and radio links can represent the more reliable solution, even if using such an arrangement means dealing with more latency.
Kalinin expounded the problems encountered with reference to a common scenario: “So [in] Ukraine, we might see footage where Ukraine FP-1 launches the FPV, and [the] FPV actually strike anti-aircraft system in Crimea. So [the] logic behind [this] is the FPV have so small signature, radar signature, the actual Pantsir system is completely blind against that. But [the] problem behind [it] is latency. On top of 600 or 300 milliseconds of latency of the Starlink, you get also radio channel frequency, so it’s like playing video games with the ping of one second. If you tried it once, you know what I’m talking about.”
“So whenever you’re trying to get like eight millisecond ping, let’s talk about 800 millisecond or one second ping. That’s exactly what we are talking about. It’s exactly illustrative example, what [the] operator and the skill to perform with this latency is the separate grade of mastery. So AI becomes [about] edge computing, which completely eliminates this problem. So you have [a] much higher efficiency of the last mile hit, and you don’t have to have even basic mastery. Because…you know auto-aim, cheats – this is exactly it. You have [a] system which always makes ‘headshots’. That’s exactly it. Even few ‘mastery’ [level] operators can do that, but in this case, you have exactly that. That’s what AI is”, Kalinin added.
The high skill level required of a drone pilot to conduct strikes using high-latency connections necessarily imposes a bottleneck on the number of such strikes which can be conducted, since very high-skill drone pilots represent only a small percentage of the overall pool. As such, AI is envisioned as a way to remove this skill barrier, thereby effectively democratising accurate long-range strike capabilities. Beyond strikes, other functions could also be suited for automation, such as reconnaissance, or remote mining of enemy territory.
However, it is effectively impossible to contain a technology once it has become commonplace, as with AI. In this vein, Kalinin warned about the concerning proliferation potential of such technologies: “[Presently] Ukraine is one step further [ahead] of Russia in this matter, but we [are] going to have Russia one step further in this. Once we have that, it’s like breaking the dam…Once we have that, [we could see] exporting FPV technology to the Hezbollah, to the Iran, to [whomever]. Right now, it’s exporting actual people – experts. They have become mercenary and come to another country…this is what it is right now. With that technology, you don’t have to actually export experts. You just export a piece of gear and they could use it efficiently.”
The proliferation of cheap AI-targeted weapons could indeed represent a major threat to militaries around the world, and other features such as optical geo-positioning, as well as spatial AI navigation, both of which already exist, could be integrated. The end result would be widely-available and low-cost, accurate, munitions that require little skill to use and are effectively immune to jamming. In the wrong hands, these would have the potential to cause seriously destabilising effects.
Warsight asked Kalinin if it were possible to simply export the software, rather than the hardware, and to this he responded that this was not really the case due to the high level of dependency on hardware-software integration. Kalinin said: “It’s combined technology, camera, edge computing, AI model, it’s all together. It’s become less flexible from the manual perspective. Because every change in gram of weight becomes some problem of learning [for] the model. But it’s still really perspective technology, I would say. So that’s the reason why we are doing our anti-AI research, I’d say.”
The rise of AI targeting naturally also generates a need to find countermeasures against it. Counter-AI is not an entirely new field, with published academic papers on adversarial techniques aimed at confusing deep neural networks (DNNs) dating back over a decade. What is new is the relevance and practical applicability of such adversarial counter-AI techniques to the battlefield.
This situation has arisen due to the explosive rise of AI since 2022, with the technology becoming increasingly common and significantly more capable and sophisticated as different models have aimed to out-compete one another within a highly competitive international market.
Kalinin described the key potential attack vectors for dealing with the kind of ‘machine vision’ employed by drones for terminal guidance: “There are…most importantly two layers of influence – one of the camera itself, but you need to know the physical [characteristics] of the camera, the specific model of the camera, because it’s parameter-specific. Another way to influence is the computing model behind, that’s [the] more advanced thing.”
“When you’re attacking the camera, you need to know the shutter speed and some specific parameters, and you have to presume, and if you didn’t guess [correctly], you will not be efficient at all. If we talk about the AI algorithm, there are a few algorithm steps which are very common for all AI algorithms, CNN [convolutional neural network] networks, et cetera.” Kalinin explained.
The types of counter-AI techniques Kalinin alluded to bear further explaining. In simple terms, they aim to find and exploit common vulnerabilities in the ways that numerous AI models process visual information. These limitations to visual processing should be familiar to anyone who has had to solve a visual CAPTCHA puzzle when accessing a website, as these work on a similar principle.
In some ways, the counter-AI field looks to have a similar problem set to cybersecurity, insofar as it becomes about finding vulnerabilities and building countermeasures around them on the one side, while the other side’s aim becomes training AI models to be less susceptible to many types of adversarial patterns.
The results of these adversarial patterns can be quite impressive. In one instance, researchers used a 3D printer to fabricate a turtle and a baseball, each of which contained precise embedded adversarial ‘perturbations’ designed to disguise them as another object. These patterns can sometimes be effectively invisible to the human eye, but are noticeable by machine vision. Images of the turtle and baseball, taken from different angles, were then fed into an image recognition algorithm, which then classified the turtle as a rifle, and the baseball as an espresso – with remarkable consistency.
In another case, researchers were able to introduce precise perturbations onto an image of a stop sign, to trick an image recognition algorithm into seeing it as a yield sign, and thereby potentially cause a self-driving car to behave dangerously. In yet another example, researchers from Keen Security Lab, a division within Chinese tech giant Tencent, conducted security research on a Tesla Model S 75. The researchers managed to trick the car into activating its windscreen wipers, and also found ways to ‘hide’ lanes from the car’s lane-detection system, and even designed a ‘fake lane attack’ to cause the car’s autosteer to move the car into the lane for oncoming traffic.
Lastly, one team of researchers were able to pass unnoticed in front of a person detector – a camera with embedded machine vision software for detecting people – simply by wearing a t-shirt with an embedded adversarial design. While the researcher wore the t-shirt, the camera did not register them human, and so did not detect them.
All of the aforementioned examples used adversarial patterns to disrupt or spoof machine vision due through common vulnerabilities in the way image recognition algorithms work. Now, similar techniques look set to be in demand on the battlefield as AI targeting takes off. In images published on 2 June 2026, Russian forces have already been seen conducting experiments with painting adversarial patterns aimed at deceiving machine vision onto their trucks.
With regards to Russia’s own adoption of AI targeting, some of their efforts have appeared in open sources since at least early 2025. While they present a cause for concern, Kalinin said there was still time to develop solutions, remarking, “We are actually interested in cooperation on anti-AI systems, because we have still have some time to countermeasure this technology.”
Russia meanwhile, has continued to conduct various experiments to gain an edge in the drone war, some of which have been covered previously. A portion of these are likely to benefit from AI targeting integration, with a case in point being Russia’s own use of mothership drones. These usually comprise a Geran-2 or Gerbera loaded with two FPV drones, controlled via a radio link bridged via the mothership.
When Warsight asked Kalinin whether the Russians were using mothership attacks mainly to engage specific, pre-planned targets, or simply using them to engage targets of opportunity en route, Kalinin said:
“Basically both, so you could go for a specific target when you deploy FPVs knowing the target, but FPVs have one big advantage in contrast to the regular anti-tank missiles for instance. For anti-tank…missiles, you see the tank, you’re reacting in minutes or seconds basically, you don’t have time to think about it. With FPV you have 15 minutes to do whatever you want. So you could see the target, choose the target, wait, just chill in the air to see the better opportunity to attack, analyse [the] target, select another target, whatever you want. So if they [are] deploying over the city it’s always target could be spotted. Worst case scenario they hit [a] civilian one…but in the regular scenario you will see really important piece of infrastructure to be hit, at least to cause the panic.”
Kalinin noted that these kinds of infrastructure targets tended to include “cellular network towers, maybe electrical transformers, maybe some gasoline stations, vehicles, military-painted vehicles, et cetera…And you have like 15 minutes to select the target, you don’t need additional recon or whatever, and the target could be high velocity it doesn’t matter at all. And it’s enough explosive to hit vehicle from target from the vehicle size onwards point of view….Trains, you could attack multiple parts of [a] train to cause more damage, you can do the ‘double tap’ with that because you have 15 minutes from the launch time, and launch time you could select.”
Kalinin also explained that on motherships “it’s always more than one FPV otherwise it doesn’t make any sense, the carrier itself could hold more explosive so…typically it’s two FPVs.” This arrangement allows the mothership to loiter in an area for an area, and conduct two attacks with FPVs, before then engaging a target itself with its own on-board warhead.
Another related problem for Ukraine has been Russia’s increasing adoption of mesh networking into its OWA UAVs such as Geran and Gerbera, which have proven challenging for Ukraine’s EW systems to deal with. According to Kalinin, the reason for their relative effectiveness at present is down to the limitations of Ukraine’s current jammer network.
Kalinin said: “It’s not about how much jammers we need, it’s about what we built and deployed for the last few years. So we built up the network of jammers over Ukraine which [is] mostly concentrated on countermeasure of navigation, so it’s all about the jamming or spoofing of GNSS [Global Navigation Satellite System] signals. It doesn’t mean it could do anything with the mesh control and command, and that’s a big problem. So [what’s] dangerous with mesh [is] not that it’s some unique durable technology, it’s all about the ability and flexibility [of the] network which [is] already deployed in the field. It doesn’t have enough flexibility so it has to be modernised and we need to find out the advanced way to countermeasure these on the vast areas of Ukraine. It’s quite a big country.”
The problem as outlined by Kalinin is that Ukraine’s jammer network is insufficiently deep, meaning that there can often be little or nothing in place to jam threats once they have passed through the established jamming layer closer to the front lines. “If you penetrate 100 kilometres from the front line, you have relatively zero jammers which [are] able to countermeasure the mesh [network]. And…by inertial system you will penetrate it anyway, so you could turn it on and one transmission hop for 100 kilometres is not a problem at all, so it’s having two systems operating separately” Kalinin explained, adding ,“the only reason why it’s so why it’s efficient – because there is no jammers on the spot.”
Russia’s increasing adoption of mesh networking has also allowed their forces to explore various new tactical approaches in drone warfare. According to Kalinin, “they actively test the swarms. They actively test recon and attack on our small manoeuvre anti-aircraft groups…and they have success recently with that. I mean, noticeable success.”
Kalinin explained how these attacks on Ukrainian mobile air defence groups worked: “So they already have installed their video cameras on their Gerans or Shaheds…It’s utilising free bandwidth of the mesh network. You could spot the position of cannons, anti-aircraft cannon system, which [are] trying to shoot down the Geran nearby. So the first one spots the position and on the second or third hop of the mesh network, [the] operator comes into picture and actually try to attack the position…in manual mode.”
Not all of Russia’s drone experiments have been so successful. In a case Warsight reported on earlier in 2026, Russian forces had been experimenting with mounting near-infrared (NIR) spotlights on top of drones, to dazzle Ukrainian interceptor drones. Kalinin flatly said “They test that, it doesn’t work.” He added, “because we have a purely infrared camera – it’s not near-infrared. So their idea was to blind near-infrared, so the security camera you could buy from China, it used [near-] infrared.”
As Kalinin explained, these kinds of techniques would only work “if you use cheap near-infrared” camera on the interceptor drone. However, as he noted, “it’s a fraction of [a] percent of how much Ukrainian drones actually use this.” Instead, Ukrainian interceptor drones tend to be equipped with thermal infrared cameras – such as medium-wave IR (MWIR) and long-wave IR (LWIR). These are immune to NIR dazzling, and indeed the spotlight can indeed make the drone even easier to spot, as Kalinin concluded, “so it basically makes our interceptor more efficient – it works [the] other way.”
Another Russian development seen as an area of concern was the increasing speeds of the Geran family, with jet engine powered models such as Geran-3, Geran-4, and Geran-5 OWA UAVs emerging, with the latter effectively being low-cost cruise missiles.
Yet Kalinin did not seem too concerned about the appearance of these faster models, stating bluntly, “jet technology in Shaheds is not a game changer”. He explained that firstly he expected Russia to operate a “mixed system” consisting of both propeller- and jet-powered OWA UAVs. He said this was in part because propeller-powered models had some advantages over jet-powered models in terms of range and payload, but also because there were likely further survivability gains for propeller-powered models to be eked out from improving their evasive manoeuvring against interceptors.
Russia has been experimenting with evasive manoeuvring, by mounting interceptor approach warning cameras on several types of drone since at least late-2024. These work by detecting an approaching Ukrainian interceptor through machine vision, and then automatically initiating evasive manoeuvres via an on-board AI. However, Kalinin said he thought Russia’s current efforts had some way to go before they reached their peak effectiveness.
Regarding Ukraine’s challenge of intercepting of Russia’s new jet-powered Geran models, Kalinin explained: “We already have different interceptors. We see and understand they could be jet propulsion, and we see the cruise missiles way before we see that [many] Shaheds. So Ukraine [is] already building up some capabilities to intercept high-speed targets, let’s say.”
Regarding which propulsion technologies Ukraine was pursuing to engage these higher-speed targets, “I see different approaches”, Kalinin said, “that’s [the] advantage of Ukraine – don’t estimate how many engineers we have!” he added, laughing.
Indeed, Ukraine has made noteworthy advancements in increasing interceptor drone speeds, with manufacturer Wild Hornets releasing footage on 3 May 2026 of one of their interceptor drones catching up to and engaging a Russian Geran-4 in a tail-chase scenario. Geran-4 is the second-fastest model in the Geran family so far, assessed by Main Directorate of Intelligence (GUR) of Ukraine as capable of reaching speeds of up to 500 km/h, and manoeuvring at speeds of 300-400 km/h. The fastest model is the Geran-5, assessed as being capable of reaching speeds of 450-600 km/h, but this may not be fast enough to keep it safe from FPV interceptors for long.
In late May 2026, Australian company Drone Pro Hub claimed to have broken the drone airspeed record with their Blackbird FPV drone recorded reaching 730 km/h with tailwind, and 685 km/h into headwind during a 10 April 2026 flight. By way of comparison, these speeds are in the same region as commercial airliners, and while the drone did not have to carry a warhead as Ukraine’s interceptors do, it nonetheless shows that there are further performance gains to be attained from iterating upon current FPV interceptor technologies.
Alongside impressive speeds, FPV drones are also capable of impressive altitude performance. A 3 January 2019 experiment by a Russian drone hobbyist successfully demonstrated that an FPV drone could climb to around 12,500 m (41,000 ft). This is above the cruising altitude of most commercial airliners. From a recent conversation with a drone developer, weaponised FPV interceptor drones with similar levels of altitude performance exist today. However, the true altitudes attainable by such models are rarely advertised by developers, as it is generally considered a fairly sensitive topic, owing to concerns about the potential risks to civil aviation.
Beyond speed and altitude, FPV drones are now also able to strike at longer ranges. In conversations I had with Ukrainians in spring of 2025, maximum FPV drone ranges of 50 km were being quoted. Now just one year later, on 26 May 2026, Serhii Sternenko, Advisor to Ukraine’s Minister of Defence posted on his X account that a Ukrainian FPV drone has successfully engaged a Russian target at a distance of 102 km without using a mothership. The speed with which drone capabilities are evolving is truly remarkable.
Production numbers are equally remarkable. Kalinin said: “About mass-scale production, our ‘housewives’ [Note: this is a sarcastic reference to an infamous March 2026 interview of Rheinmetall CEO Armin Papperger in The Atlantic, in which Papperger claimed Ukraine’s drones were being produced by “housewives” – a comment which drew considerable mockery from many Ukrainians] are producing, I believe, around maybe more than 100,000, 200,000 interceptors per month. So we are doing some mass-scale production. There are limitations – which is [a] huge obstacle – to create the interceptor factory, like a proper way of mass scale production, which could make it two million per month. But we have multiple producers, they have multiple facilities in Ukraine, outside of Ukraine, and combining all this together, I believe, we could scale up to two million per month. We don’t need that much.”
The rise of AI targeting should be a cause for concern, because it promises to democratise mass precision strike capabilities, even at very long ranges. However, perhaps more concerning is the time in which it has arisen. At this moment in time, the world is undergoing a confluence of several important factors.
For starters, AI is fast becoming sufficiently mature for practical battlefield use, and sophisticated AI models are widely commercially available. These two factors alone will make it very difficult to prevent the threat from proliferating. Adding to these is that robust, GNSS-free navigation is also becoming increasingly available, making the task of defenders much more difficult, since such technologies can allow attackers to effectively bypass common defences such as GNSS and control channel jamming.
Additionally, FPV drone performance in terms of speed, range and altitude is reaching the point where many kinds of aircraft may now realistically be targeted – especially when combined with the potential for range extension from carrier motherships. At the same time, FPV drones can be produced at relatively low cost, and in massive quantities. All of this comes at precisely the time when no military in the world really has a decisive answer to the problem of scale.
To put the problem of scale into perspective, another speaker at AOC Europe 2026 was Colonel Ivan Pavlenko, Chief of Ukraine’s General Directorate of Electronic and Cyber Warfare. Pavlenko stated, “Today, UAVs totally dominate in the air and they ensure no less than 70% of all enemy losses”. According to Pavlenko, “every day [the] Russians launch no less than 7,000 FPVs, 1,500 winged-type kamikaze drones, and provide no less than 2,000 surveillance flights.”
Pavlenko also bluntly stated that with the rise of drones on the battlefield “[the] world will never be the same. We have a whole line of different drones. Kamikaze drones, surveillance, interceptors, logistics. Today drones fight with drones, provide logistics, and even evacuate casualties. [The] ‘kill zone’ is a new existing reality…today, drones totally control the combat area.”
Militaries the world over have demonstrated numerous solutions to take down drones, but in most cases this has been done at relatively small scales, and often using a mix of jamming and hard-kill solutions. When the problem set is expanded to include thousands of targets every single day, the challenge increases by orders of magnitude. Looking ahead, when unjammable AI navigation and targeting is layered on top of this mass, along with impressive speed and altitude capabilities, the use of motherships to remove previous range limits, and distributed mass production, the problem becomes something expected to keep military planners awake at night.
Yet this is precisely the world we are rapidly entering. The sheer speed with which these battlefield adaptations are happening is truly stunning. Many of the limits that existed just one year ago are already disappearing. At present, the consensus seems to be that AI on the battlefield will take time to perfect. Col Pavlenko said, “combat reality today – [it’s] not always so good with AI, and I think that it requires time and maybe years to total application of AI in our combat reality. It’s not a single process.”
Indeed, drones with AI targeting or navigation currently represent a minority of the drones in use today. However this could shift very rapidly. Consider that in early 2025, interceptor FPV drones were only just coming onto the scene. Now, merely one year later they are being produced by the hundreds of thousands per month, capable of attaining greater ranges and higher speeds than before, and are in the process of integrating AI targeting.
Looking ahead, while Ukraine’s current lead in AI targeting gives their forces an edge, it is worth remembering Kalinin’s warnings. Technology is always a double-edged sword, and while Ukraine may be ahead today, Russia is also exploring the applications of AI integration, and may pull ahead tomorrow. Furthermore, once fielded, such technologies are unlikely to remain within the confines of Russia’s or Ukraine’s borders – others will have uses for what these technologies can provide.
Both Ukrainian and Russian forces have managed to continue operating in the face of massed drone strikes, but for both sides this resilience has often relied upon their respective populations’ capacity to absorb casualties by the many tens of thousands. How many militaries could, or indeed would want to go the same route?
Pandora’s box has been opened, and closing it may well prove impossible. However, as in the ancient Greek tale, counter-AI techniques may, for the time being at least, represent the final thing left in the mythical box – hope.
