Russian Military AI: Moscow's Claims vs. the Evidence

Russian military AI is in the news again, and so is a sharp claim. On October 12, 2026, Russia's defence ministry said it had opened a service that lets drone makers train AI on recorded combat footage. The same day, a Russian analyst told Sputnik the push answers NATO's AI tests in Ukraine, and that Palantir and Anthropic had supplied their systems to Kyiv.

Some of that holds up. Some of it doesn't. We checked each claim against public records: what Russia announced, what's known about the two US companies, how much these systems decide on their own, and why the real race is about data.

Key Takeaways

• Russia's defence ministry opened a combat-footage service for training military AI on October 12, 2026. Sputnik says more than 40 groups have joined.
• A Russian analyst, Alexei Leonkov, told Sputnik that Palantir and Anthropic supplied their systems to Ukraine. He gave no evidence.
• The Palantir part is on the record. Its software has helped Ukraine's forces since 2022, it opened an office in Kyiv, and in January 2026 it co-built the Brave1 Dataroom for training AI on real battle data.
• We found no public evidence that Anthropic supplied Claude to Ukraine's military. Its own September report shows Russia-linked groups misusing Claude against Ukraine.
• Both sides describe military AI mainly as a tool that helps people decide. The race underneath is about who has the best labelled data.

Russian Military AI at a Glance

• What's new: a defence ministry service, launched October 12, that shares drone combat footage with developers.
• Who runs it: Deputy Defence Minister Dmitry Shcherbinin, per Sputnik, with the ministry checking every upload and then using the pooled footage to train its own models too.
• Scale: more than 40 groups. That's Russia's own figure, and so far nobody outside Russia has been able to check it or say how much video the library holds.
• Command system: Svod, a tool that merges battlefield data into one map. Russia says it's partly in use.
• The claim: that Palantir and Anthropic gave their systems to Ukraine.
• Sources: Sputnik and TASS, both Russian state media, plus Western research and company reports.

What Did Russia Announce?

The news is a data service, not a new weapon. TASS says the ministry built a library of drone combat video. It covers FPV drones, other rotor drones and fixed-wing drones, and drone makers can draw on it to train and test their own models before those models ever fly.

The service has three parts:

• A footage library of real combat use, locked behind certified encryption and access checks, the ministry says.
• One labelling scheme. Ministry staff built it so every developer tags data the same way.
• A feedback loop. Developers upload their results, the ministry checks the format, then it uses the pooled data to train its own models.

Sputnik adds that more than 40 groups have joined. Some, it says, already put trained models into their combat drones. TASS ties the effort to a plan Defence Minister Andrey Belousov set out in July 2026.

Treat these numbers as Russia's own. Both TASS and Sputnik are state media, and neither gives any way for an outsider to check the 40-group figure or how much footage the library holds.

What Did the Russian Analyst Claim?

The claim came in a second Sputnik piece the same day. Alexei Leonkov, a military analyst, called Russia's work a tit-for-tat answer to NATO's AI tests in Ukraine. Outlets in Vietnam and elsewhere ran the story a day later.

His main points, as Sputnik reported them:

• Palantir and Anthropic have supplied their systems, and Ukraine is using them in combat, not just testing them.
• These tools give Ukraine an edge, mostly in how fast commanders decide.
• Russia wants AI that turns a flood of reports, video and sensor feeds into one clear picture, because, he said, even the best commanders can't sort it all fast enough on their own.
• Russia has done this before, with the Svod system.
• It would be wrong to call Russian military AI fully autonomous.

He also said the other side is testing these systems against Iran. Sputnik gave no proof for that, or for the claim about the two companies.

Where Is Russian Military AI Used Today?

Leonkov's list is worth reading on its own, because it shows where Russia says it wants AI to help. These are his claims as Sputnik reported them, not facts that we or anyone outside Russia could confirm, so read the list as a statement of intent rather than a record of what works.

• Sorting data. Pulling reports, drone video and sensor feeds into one live picture that a commander can act on without waiting for staff to piece it together by hand.
• Spotting and sorting targets. Then passing threat data to air, coastal and ground defence.
• Air defence. Some missile systems can fire on their own, he said, and the idea could spread to planes and ships.
• Reading new enemy weapons. Working out what a newly seen enemy system can do, and how to counter it, in a few hours or even minutes, according to Leonkov.
• Planning big campaigns. Strikes on depots, ports and rail yards, and defence against large drone raids.

Most of this is about speed. The same tasks took staff officers days in past wars. Leonkov's point is that AI can shrink that to minutes, while people still sign off.

Is the Palantir Claim True?

Mostly, yes. Palantir's role in Ukraine is public, it goes back years, and the firm has talked about it openly.

Here is what's on the record:

• Since 2022: Palantir has worked with Ukraine's government. It opened a Kyiv office and works with the Ministry of Digital Transformation.
• MetaConstellation: its software pulls in data from satellites, sensors and drones to help Ukraine map enemy positions, Defense News reports.
• The CEO's own words: at Davos in January 2023, Alex Karp said Palantir's tools had made Ukraine's targeting far better, as DefenseScoop reported.
• Brave1 Dataroom: in January 2026, Ukraine's Brave1 defence tech cluster and Palantir opened a secure space. Approved Ukrainian defence firms use it to train and test AI on real, sensitive data about Russian air threats, starting with Shahed-type attack drones.

So on Palantir, Leonkov is right, though there's nothing secret or new about it: the company has talked about its Ukraine work in public for more than three years.

Is the Anthropic Claim True?

We found no public evidence that Anthropic gave Claude to Ukraine's military. Leonkov named no product, deal, unit or date. We found no Ukrainian, US or company source that describes one.

Claude is sold in many countries, so we can't rule out that someone in Ukraine's defence sector uses it. But "supplied their systems" for combat is a specific claim. It needs more than one analyst's word.

Here is what is on the record about Anthropic and defence:

• Its stated view: in a February 26, 2026 statement, CEO Dario Amodei wrote that partially autonomous weapons, "like those used today in Ukraine," are vital to the defence of democracy. He also said today's AI isn't reliable enough for fully autonomous weapons.
• Its US work: the same statement says Claude was widely used across the US Department of War and other security agencies. Seoul Economic Daily, citing the BBC, reports it was built into Maven Smart System, a data platform run by Palantir.
• Its fight with the Pentagon: Anthropic refused to drop its limits on fully autonomous weapons and mass domestic surveillance. The Pentagon then labelled it a supply chain risk. In early October, it said it had stopped using Anthropic's products, according to reports.

That US link may explain why the two names appear together. Palantir and Anthropic did team up, in a deal announced in November 2024, but for American agencies. We found nothing in any public source that ties that work, or any other Anthropic deal, to Ukraine's armed forces or to NATO tests on Ukrainian soil.

What Does Anthropic's Own Report Say About Russia?

The most detailed public record points the other way. Anthropic's September 2026 threat report covers misuse from December 2025 to August 2026. Several cases involve Russia-linked actors aiming at Ukraine.

The two cases that matter most here:

• GTG-20006: a spying group that used Claude Code against more than 20 organisations. Most were Ukrainian and European government, military and diplomatic bodies, plus drone suppliers. Anthropic links it to Midnight Blizzard, based on public reporting.
• GTG-27005: a small freelance team in Russia that used Claude Code to write software for a drone swarm. Its members trained a target classifier on scraped Ukrainian combat video, sorting what they saw into friend and foe, and used a fixed point in Donetsk region as their demo target. The design let the drones strike with no human in the loop.

Anthropic rated the drone work as early-stage, tested only in simulation. It found nine linked accounts and banned them. Eight were used only for ordinary freelance jobs. Anthropic said the team wasn't a state body, but had ties to a regional university linked to the Russian Academy of Sciences.

One detail matters for anyone who runs an AI service. Claude isn't offered in Russia. The actors got around the block by renting servers abroad. Our article on Anthropic's usage policy update covers how its rules on weapons and drones have changed since.

What Is Svod?

Svod is Russia's tactical awareness system. Defence Minister Belousov announced it in August 2025, and Leonkov now says it's partly in use.

Reports describe it as software, not a device, that runs on existing military laptops and tablets and merges satellite, drone, spy and open-source data into one map. Some outlets say it uses AI to suggest courses of action. Others stress that it can't decide on its own.

A CSIS analysis from February 2026 is cautious. Its author, Kateryna Bondar, notes it's unclear whether Svod is truly new or a rebrand of older projects that missed their goals. The real test, she argues, is whether decisions get faster in the field. Official claims don't prove that.

How Autonomous Is Russian Military AI?

Less than the headlines suggest, by Russia's own telling. Leonkov said people still have to weigh what AI can't, such as the risk to civilians and to homes, power and water.

He did say some missile systems can fire on their own. That's not new. Defences built to stop incoming missiles have had automatic modes for decades, because people can't react fast enough.

The bigger worry sits outside official programmes. The GTG-27005 case shows a small freelance team trying to build a swarm that picks its own targets, people included. No human would approve each strike. That's the line Anthropic says it won't cross, and one that many rights groups want banned.

Ukraine is adding autonomy too, mostly for defence. The Brave1 Dataroom's first goal is to help interceptor drones find and stop Shahed-type drones. They come in numbers no human team can track one by one.

Russian Military AI Claims vs. the Evidence

Claim	Source	What we found
Russia opened a combat-footage service for AI training	Russian defence ministry, via TASS and Sputnik	Announced October 12, 2026. Not checked outside Russia.
More than 40 groups have joined	Shcherbinin, via Sputnik	Russia's own figure. No outside check.
Palantir supplied systems to Ukraine	Leonkov, via Sputnik	On the record since 2022, from MetaConstellation mapping software to the Brave1 Dataroom that Palantir co-built with Ukraine in January 2026.
Anthropic supplied systems to Ukraine	Leonkov, via Sputnik	No public evidence found. Anthropic's report shows Russia-linked misuse of Claude against Ukraine.
Ukraine tests these systems against Iran	Leonkov, via Sputnik	No evidence given or found.
Svod is partly in use	Leonkov and Russian officials	Plausible but not confirmed. CSIS asks whether it's truly new or a rebrand of older Russian command projects that never met their goals.
Russian military AI is not fully autonomous	Leonkov, via Sputnik	Matches the CSIS view. Some air defence fires on its own.

Why Are Both Sides Building Data Rooms?

Because the models are the easy part. Footage, labels and feedback are what make military AI useful. Both sides have now built a shared place to pool them.

Look at how alike the two set-ups are:

• Russia's service offers a combat video library, one labelling standard, and a loop where developers send results back.
• Ukraine's Brave1 Dataroom offers curated visual and thermal data on aerial targets, access for vetted firms, and a shared place to train and test.

The goal is the same in both. Many small developers train on one trusted dataset instead of each building their own. The state keeps the combined result.

That's a familiar pattern in civilian data work. One shared, well-labelled dataset with clear access rules beats a pile of private ones almost every time, whether you're spotting drones, fraud or faulty parts on a factory line.

What This Means for Companies That Build AI

Few teams reading this will ever touch defence work. Still, the lessons carry over almost one for one to any company that trains models on its own data:

• Set one labelling standard early. Russia built its classifier before it opened the doors. Mixed labels make pooled data hard to use.
• Close the feedback loop. Results that flow back into one dataset improve every model trained on it.
• Keep a person accountable for high-stakes outputs. Even here, officials on both sides insist a human makes the final call. Our guide to human review loops shows how to set that up.
• Expect misuse. If you offer an AI tool, people you'd never sell to will try to reach it. Country blocks alone won't stop them, so pair them with the controls in our agentic AI security guide.
• Watch the policy risk. Anthropic's fight with the Pentagon shows a vendor's rules can decide who may use a model. Write it into your AI governance plan.

Data quality is where most AI projects win or fail, in war or in retail. If you're setting up a labelling standard or a shared training dataset, that's the work we do.

How to Read Russian Military AI Claims

Stories like this mix fact, spin and guesswork, often in the same paragraph, and they spread fast because few readers have time to trace each claim back to where it started. A short checklist helps:

• Who is talking? State media on any side has a reason to shape the story.
• Is there a primary source? A contract, a company report or an official statement beats an analyst's comment.
• Can you check the numbers? "More than 40 groups" is a claim until someone outside confirms it.
• What's the date? Old quotes, like Karp's from 2023, often come back as if they were new.
• Who gains if you believe it? That includes us, which is why our position is stated at the top.

What Is Still Unclear?

Several questions are still open:

• How large and how good Russia's footage library really is.
• Whether Svod is making Russian units decide faster.
• What any Ukrainian or NATO system actually runs under the hood.
• How the Anthropic and Pentagon case ends after the appeals court ruling, and whether other AI vendors now write similar limits into their own government contracts.

We'll update this page if any of these get clearer answers.

What the Russian Military AI Race Means Next

The Russian military AI race has moved from showcase drones to data pipelines, and both sides now say so openly. Expect more shared datasets, more automation in air defence, and more fights over where human control must stay.

The claim against Anthropic, as it stands, has no public evidence behind it. The one about Palantir is old news. The real story is the quiet one: whoever builds the best labelled data, fastest, gets the edge.

For more on how AI companies police their models, see our breakdown of Anthropic's distillation report.