Anthropic says it disrupted a Russia-based operation to engineer an autonomous kamikaze drone swarm using Claude

Anthropic disclosed that it thwarted a Russian operation aiming to create an autonomous drone swarm, utilizing Claude, for kamikaze drone attacks. This intelligence report, part of Anthropic's routine assessments on curbing AI-related malicious activities, detailed the group's utilization of Claude Code, a coding tool, to script and assess software for their project titled "DronDoc" or "Serafim."

Additionally, they employed a software simulation suite and a graphics processing unit-equipped server for model training. With AI assistance, the group established shared swarm memory, fault-tolerant coordination logic, an onboard small language model for directing drone actions, a guidance software system to navigate drones towards targets, and a geolocation module to identify opposing drone operators.

The system also incorporated a passive acoustic detection layer and low-level logic for the drones' programmable chips. The platform's intent was for autonomous lethal applications. The developers purportedly crafted the platform to independently target "persons," including enemy combatants, without human intervention. They trained a computer-vision classifier using Ukrainian battlefield footage, categorizing Russian systems as friendly, and repeatedly used specific coordinates in Ukraine's Donetsk Region as a target.

The operation began in late 2025 and intensified in mid-May of the following year, involving a small, freelance group from a regional university and a federal research center linked to the Russian Academy of Sciences. The developers claimed Russian-government funding, though Anthropic could not independently verify this. In response, Anthropic blocked these accounts and integrated the findings into their safeguards to mitigate future risks.