Russian developers used Anthropic’s Claude to build software for an autonomous kamikaze drone swarm that can select targets without human approval.
According to Anthropic the group appeared to be a small freelance team rather than a Russian state organisation, and had used Claude Code to develop a system called “DronDoc” or “Serafim”. The project included swarm coordination, terminal guidance, acoustic detection and low-level firmware, along with simulation tools and rented GPU capacity for training.
Claude helped the developers create software for attack, observation and return-to-base behaviour using an onboard small language model. Anthropic said the guidance system could steer drones towards targets using onboard cameras and issue a detonation command without human intervention.
“The onboard model could select targets (including a ‘person’ target class) and issue detonation commands without a human in the loop,” Anthropic said.
The developers trained a computer vision classifier using scraped Ukrainian combat footage and divided targets into “enemy” and “friendly” categories. Russian systems were placed on an allow-list, while locations in Donetsk Oblast and other Ukrainian front-line areas appeared in the project’s mission planning.
Anthropic said the group had progressed beyond simulation, flashing firmware onto real development boards and linking systems across a mesh network. The company rated parts of the project at Technology Readiness Level three to four, covering proof-of-concept and simulation validation rather than operational deployment.
Anthropic linked nine accounts to the group and said the operation began in May 2026, with developers using commercial virtual private servers to bypass geographic restrictions. Anthropic banned the accounts after detection and said the case helped strengthen safeguards against AI-assisted weapons development.
