In the latest edition of its threat intelligence report, AI corporation Anthropic said it had disrupted the activities of a Russian group that was using Claude to try to develop a fully autonomous swarm of FPV kamikaze drones. The September edition is part of Anthropic’s series of regular reports on efforts to counter malicious uses of artificial intelligence.
According to the company, the developers used Claude Code, Anthropic’s AI coding tool, to write and test software, saving the results directly into files for a project they called “DronDoc” or “Serafim.”
Apart from Claude, the developers used a software simulation stack and a rented server equipped with graphics processing units to train models. With the help of AI, the group developed:
- shared swarm memory and fault-tolerant coordination logic (FTCL)
- an onboard small language model to govern attack, observe, and return-to-base behaviors
- a terminal guidance software system to steer drones to their target (using the onboard camera) and issue the call to detonate
- a control-link geolocation module to find opposing drone operators
- a passive acoustic detection layer
- and low-level logic for the drones’ programmable chips
The platform was designed for autonomous lethal use. Anthropic said the onboard model could independently select targets — including a “person” target class — and issue a command to detonate without human involvement.
Using captured battlefield footage from Ukraine, the developers trained a computer-vision classifier that divided targets into “friendly” and “enemy” categories, placing Russian systems on the friendly list.
They repeatedly used the same coordinates in Ukraine’s Donetsk Region as a demonstration strike point, while the mission area itself covered a wider range of front-line cities and logistics corridors inside Ukraine.
Accounts linked to the operation were created between late 2025 and early 2026, and the activity itself began in mid-May of this year. The developers used commercial virtual private servers, or VPS services, to circumvent geographic restrictions on access.
Anthropic concluded that the people behind the project were a small, specialized freelance group handling both civilian and military contracts, rather than an organization directly tied to the Russian government.
Of nine accounts linked to the group, eight were used only for ordinary freelance work unrelated to weapons.
According to Anthropic, members of the operation appeared to have ties to a regional university and a federal research center affiliated with the Russian Academy of Sciences.
The developers themselves claimed they were receiving funding from Russia’s Advanced Research Foundation, a state defense research body sometimes compared to the U.S. Defense Advanced Research Projects Agency (DARPA), as well as the National Technology Initiative and the Russian Defense Ministry. Anthropic said it could not independently verify those claims.
The company blocked accounts associated with the group and incorporated findings from the investigation into its safeguards in an effort to prevent similar abuses in the future.
In November 2025, Anthropic reported that Claude had been used to almost completely automate a major cyberespionage campaign, with 80% to 90% of the activity carried out without human involvement. Analysts attributed that operation to GTG-1002, a group linked to the Chinese government. It targeted around 30 organizations worldwide, including technology companies and government agencies.
