In brief: Anthropic, the AI safety and research company said that it had prevented Russian developers from using its Claude coding agent to create software to power the Kremlin’s autonomous kamikaze drones. The system was designed to select and attack targets, including individuals, without the need for operator intervention.

Sept. 14, 2026

As concerns grow around the direction that unfettered Artificial Intelligence development may be taking, a report by the AI safety and research company Anthropic underlines the need for oversight of the use and development of AI technology if the doomsday scenario is to be avoided.

Who used the Claude coding agent to develop killer drones?

In its quarterly report “Detecting and Countering Misuse of AI,” published earlier this month, the company identified a group of Russia-linked developers, with the designation GTG-27005, that was using its Claude coding agent to create software for autonomous kamikaze drones.

Anthropic believes that, although the system was being developed for military use, the group’s members were likely independent, university-based specialists. It had not identified any specific links with the Russian Armed Forces as the group used a number of virtual servers to avoid geographic restrictions. It said the group showed qualities consistent with the Russian hacking group “Midnight Blizzard,” which had used generative AI tools to prepare cyberattacks. The move to the development of autonomous combat systems was a new departure for the group.

What were the developers hoping to achieve?

The Anthropic report said the operation was known as “DroneDoc” and “Seraphim.” The group registered its accounts in late 2025 but began in earnest in mid-May 2026. It used the Claude Coding agent to write and review software code and to store the resulting components directly in their project files.

The code was integrated with drone simulation software and computing resource training models rented from commercial suppliers. The resulting system was designed to select targets, including people, and issue attack commands without further operator intervention.

The ultimate aim seems to be to operate attack drones as a swarm and to make decisions autonomously. The use of Claude would allow the development of shared memory for the swarm in a fault-tolerant coordination system – a distributed network architecture that maintains synchronization, and task allocation even when individual drones or components fail and communications were lost or individual components failed.

The team was also developing an onboard compact language model that would determine the drone’s next action after swarm members engage a target: attack another target, continue surveillance, or return to base.

Trained using real combat footage from Ukraine

The developers used combat footage from the fighting in Ukraine gained through social media to train a computer vision system, in which Ukrainian personnel and equipment was designated as hostile, while Russian military systems were added to a list of targets that could not be attacked.

Demonstration simulations were carried out in Ukraine’s Donetsk region Ukrainian frontline cities and transportation corridors.

Anthropic battles the illicit use of AI in the war in Ukraine

In total, the company identified and subsequently blocked nine accounts linked to the project. Along with a regional university Anthropic also found connections with the Russian Academy of Sciences.

The report also describes the growing use of AI to organize cyberattacks against Ukrainian government officials and military targets. One Russia-linked group used AI in phishing campaigns, malware distribution, WhatsApp account takeovers, and attacks conducted through compromised hotel Wi-Fi networks.

The slogan of the US National Rifle Association (NRA): “Guns don’t kill people, people kill people,” could just as easily apply to AI, based on this latest incident. It underlines that calls for vigilance over the development of AI are not overstated but it is the user rather than the technology that needs to be controlled.

https://www.kyivpost.com/post/84496