Russians Used Claude to Develop Software for Drone Swarms — Anthropic

Anthropic identified a group of Russian developers who used Claude AI to create software for an autonomous swarm of strike drones.

Anthropic’s report details the misuse of the company’s artificial intelligence models.

According to Anthropic, the group, designated GTG-27005, used Claude Code to write and test software for UAVs.

The operation, known as DronDoc or Serafim, began around mid-May 2026. To bypass Anthropic’s geographic restrictions, the developers used commercial virtual servers.

Claude was used to develop a coordination system for the drones, allowing them to exchange information and continue to perform their mission if individual drones were lost. The developers also created a compact language model designed to determine the drone’s subsequent behavior — whether to attack a target, conduct surveillance, or return to base.

According to Anthropic, the system developed by the group could autonomously select targets and issue commands to strike them without operator intervention.

One of the target categories specified in the software was “human.” Videos of combat operations in Ukraine were used to train the computer vision algorithms, with objects in the footage classified as “enemy” or “friendly.”

The developers’ materials also repeatedly used locations in Ukraine’s Donetsk region as demonstration strike zones, while Ukrainian frontline cities and transport routes were used as the geography for potential missions.

According to Anthropic, the group was likely a small specialized team of freelancers carrying out both civilian and military tasks, rather than a direct unit of a Russian government agency. The company identified nine linked accounts and established the developers’ connections to a regional university and a federal research center affiliated with the Russian Academy of Sciences.

The project participants themselves claimed that they had received funding from Russia’s Foundation for Advanced Research Projects, the National Technology Initiative, and the Ministry of Defence of the Russian Federation. Anthropic emphasized that it was unable to verify these claims. After detecting the activity, the company blocked the associated accounts.

Separately, Anthropic identified the use of its models by Russia-related groups to prepare cyberattacks against Ukrainian government officials and military personnel. According to the company, some of the detected activity shares characteristics with previously described operations linked to the Russian hacking group Midnight Blizzard.

In July, the Ukrainian Defense Forces successfully intercepted Russian Molniya UAVs equipped with artificial intelligence for the first time. According to preliminary data, the use of such AI-equipped drones is carried out by Russia’s 50th Varyag Unmanned Systems Brigade, which operates together with the Rubicon Center. The two formations are currently among the main operators of new drone technologies used by Russian occupation forces in this area.

In June 2025, experts from DIU examined the internal components of the Russian V2U drone. During the inspection of the strike UAV, which extensively uses artificial intelligence software, specialists found that the majority of its components were of Chinese origin.

In June, the U.S. government, citing national security authorities, issued an export control directive that suspended access to the Fable 5 and Mythos 5 AI models for all foreign nationals.