Anthropic published a 154-page threat intelligence report on September 11, 2026, detailing the misuse of its Claude Code AI assistant. The document describes how a group of freelance developers in Russia created software for autonomous kamikaze drones.
The project, tracked by Anthropic as GTG-27005 and internally named DronDoc or Serafim, began in mid-May 2026. The software was designed to identify targets, distinguish between friendly and enemy forces, and execute autonomous detonation without human intervention.
Developers utilized computer vision classifiers trained on combat footage from the Donetsk region to perform target selection and terminal guidance. To access the AI tool, they bypassed geographic restrictions using VPNs and commercial servers.
Anthropic identified the developers as freelancers with links to a regional university and a federal research center. The report also linked these efforts to broader cyber espionage activities, including data theft and attacks against Ukrainian entities.
The company confirmed that the drone system remained in the simulation and testing phases, with no evidence of operational deployment. The report highlights how publicly available conflict documentation can be repurposed as training data for weapons systems.
