Source: The Conversation – Canada
A new threat intelligence report from the AI company Anthropic has revealed that a Russian team scraped Ukrainian combat videos to train an autonomous “kamikaze” drone to choose its own targets and destroy them.
Anthropic’s report was published on Sept. 10, two days before CEO and co-founder of the company, Dario Amodei, published an essay arguing that humanity must urgently “slow the pace” of AI development.
The report is a 154-page account of the attempted misuse of Claude AI models for “malicious activity” between December 2025 and August 2026. It covers seven areas: cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development and distillation.
One case, tracked as GTG-27005, describes a small, likely freelance team in Russia that used Claude’s coding tool to help build software for a swarm of autonomous first-person-view (FPV) “kamikaze” drones — the cheap, camera-guided munitions that have become one of the defining weapons of the war.
Read more:
Ukraine’s drone strikes are having an impact on Russia — but Russian leaders remain committed to war
This should worry anyone who has ever filmed, shared or trusted footage from the war in Ukraine — one of the most visually documented conflicts in history.
It seems that any visual footage — even that created by journalists, soldiers, civilians and researchers to update the outside world and hold nations accountable — can now feed back into a conflict as raw material for weapons development and deployment on the battlefield.
Autonomous drone, can target a person
Anthropic’s report states the company’s threat intelligence teams have “investigated and disrupted multiple threat actors” who used Claude models to develop software for either weapons design and development or for intelligence-gathering and procurement for weapons programs.
The report documents six cases: three in China, two in Russia and one in Yemen.
The “likely freelance Russia-based” developers in case GTG-27005 called their project “DronDoc” or “Serafim.” They began work in May 2026 and routed their traffic through commercial servers to get around Anthropic’s location-based restrictions.
The software they built could enable a drone to pick its own targets, including targets in the category “person,” and issue the command to detonate — without a human confirming the strike first. That detail made headlines and it deserved the publicity.
However, one line in the report has received far less attention: the developers trained their targeting software on scraped Ukrainian combat footage, sorted into “enemy” and “friendly,” and tested it against a simulated target location in the region of the Donetsk oblast.
Scraped Ukrainian combat footage
The report does not specify where the footage came from. It could have come from Telegram channels used by military personnel, official and unofficial drone strike compilation accounts, open-source intelligence (OSINT) archives or open databases assembled by journalists and researchers. At this time, we simply don’t know which is the likely source.
The Centre for Information Resilience’s Eyes on Russia Map, for example, has verified and mapped more than 30,000 individual incidents using footage from volunteers, soldiers and civilians.
Drone footage and the aftermath of missile strikes are verified and geolocated by researchers who hope that they’ll be used as evidence for journalistic endeavours, war crimes investigations or simply to show the reality of the conflict in an era when generative AI can skew the narrative.
Those producing this footage are, for the most part, trying to make the war more visible and those responsible more accountable for their actions.
Documented for evidence, used for conflict
What this case shows is that footage built for one purpose can be repackaged for a very different one without changing at all. A clip filmed to document a strike and one that is scraped to teach a weapon what a human target looks like are, frame for frame, the same file.
Nothing about the image itself signals what use it will be put to. What determines that is who can access it and what they choose to do with it.
Ukraine’s own drone video system, OCHI, has amassed well over two million hours of footage since 2022, making it the largest publicly known collection of live-fire combat video globally.
A dataset of that scale is a gold mine for machine learning. However, this is not unusual when it comes to modern, digital warfare. A company in the United States, Enabled Intelligence, states that they recently packaged more than half a million hours of Ukrainian drone footage into a commercial dataset which was sold for military AI training.
Indeed, this kind of extraction is becoming routine and openly discussed. GTG-27005 demonstrates that it isn’t confined to declared, sanctioned users.
A reverse-engineered vision system
Read more:
Ukraine’s drone strikes are having an impact on Russia — but Russian leaders remain committed to war
This isn’t an isolated case, either. The same report describes a separate incident, tracked as GTG-20006, and connected by Anthropic to the Russian state-linked group publicly known as Midnight Blizzard.
In that case, operators stole a software kit for a military drone’s vision system and spent days using Claude to reverse-engineer it — recovering the design, components and details of an unreleased product.
Two separate groups within the same period went after the same kind of target: the system that allows a drone to recognize what it sees. Is this the beginning of a pattern — one where a drone’s ability to visually identify a target is becoming just as valuable as its engine or explosive payload?
Who else is watching?
None of this is an argument against documenting conflict. Far from it. Those visuals remain one of the few tools that ordinary people have to hold power accountable during moments of strife.
However, it is a reason to take seriously that those producing this record — journalists, soldiers, civilians volunteer researchers — are part of a supply chain whose subsequent uses they cannot predict and never agreed to.
Most conversations about the ethics of war footage have focused on who should watch it and how. This case suggests that we also need to ask who else is watching and what are they going to do with it.
John R. Heckman does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.
Original source: https://analysis1.mil-osi.com/2026/09/15/russian-team-misused-claude-ai-to-train-a-killer-drone-on-scraped-ukrainian-war-footage/