A new threat intelligence report from Anthropic, the AI company behind Claude, has uncovered a Russian effort to build autonomous "kamikaze" drones using the company's technology, with the targeting software trained on scraped Ukrainian combat footage. The report, published on September 10, details how a small, likely freelance team in Russia developed software for first-person-view (FPV) drones capable of selecting their own targets—including individuals—and initiating strikes without human confirmation.
The case, tracked as GTG-27005, is one of six documented instances where Claude models were used for malicious purposes, including weapons development and intelligence gathering. Anthropic's threat intelligence teams have been actively investigating and disrupting such actors, with cases spanning China, Russia, and Yemen.
Training on battlefield footage
According to the report, the Russian developers, who called their project "DronDoc" or "Serafim," began work in May 2026. They routed their traffic through commercial servers to bypass Anthropic's location-based restrictions. The software they built could enable a drone to autonomously identify targets, including those categorized as "person," and issue detonation commands without human oversight.
What has drawn less attention is that the developers trained their targeting software on scraped Ukrainian combat footage, sorted into "enemy" and "friendly" categories, and tested it against a simulated target location in the Donetsk oblast. The report does not specify the exact source of the footage, which could range from Telegram channels used by military personnel to open-source intelligence archives or databases compiled by journalists and researchers.
This development raises profound concerns about the dual-use nature of visual documentation. Footage captured to document war crimes, hold nations accountable, or provide journalistic evidence can now be repurposed as training data for autonomous weapons. As the report notes, a clip filmed to document a strike and one scraped to teach a weapon what a human target looks like are, frame for frame, the same file.
Ukraine's own drone video system, OCHI, has amassed over two million hours of footage since 2022, making it the largest publicly known collection of live-fire combat video globally. Such datasets are a gold mine for machine learning, and this case demonstrates that extraction is not confined to sanctioned users. A US company, Enabled Intelligence, has already packaged over half a million hours of Ukrainian drone footage into a commercial dataset for military AI training.
A broader pattern
This is not an isolated incident. The same report describes another case, GTG-20006, linked to the Russian state-sponsored group Midnight Blizzard. In that instance, operators stole a software kit for a military drone's vision system and used Claude to reverse-engineer it, recovering design and component details of an unreleased product.
Two separate groups targeting drone vision systems within the same period suggests a growing focus on the ability of drones to visually identify targets—a capability as valuable as the engine or explosive payload. This trend aligns with broader developments in the region, such as China's multidomain drone swarms and stealthy unmanned bombers, indicating that autonomous systems are becoming central to modern warfare.
Anthropic's report also comes amid growing concerns about AI's role in conflict. CEO Dario Amodei published an essay two days after the report, arguing that humanity must urgently "slow the pace" of AI development. The report itself is a 154-page account of attempted misuse between December 2025 and August 2026, covering cyber operations, influence operations, surveillance, scams, biological misuse, and conventional weapons development.
For those who have filmed, shared, or trusted footage from the war in Ukraine—one of the most visually documented conflicts in history—this case is a stark reminder that such material can feed back into the conflict as raw material for weapons development. The Center for Information Resilience's Eyes on Russia Map, for example, has verified and mapped more than 30,000 incidents using footage from volunteers, soldiers, and civilians, all intended for accountability and evidence.
While this does not argue against documenting conflict—those visuals remain crucial for holding power accountable—it underscores the need for vigilance about how such data can be exploited. As AI continues to evolve, the line between documentation and weaponization becomes increasingly blurred, with implications for Japan's drone surge and other regional efforts to develop autonomous capabilities.
