Anthropic identified a group of Russia-based freelance developers who utilized the Claude AI model to create software for an autonomous combat drone swarm, known as DronDoc or Serafim. The software facilitated swarm coordination, computer vision, and terminal guidance, allowing drones to identify targets and execute detonation commands without human intervention.
The developers trained their computer vision systems using combat footage from Ukraine and conducted simulated missions based on Ukrainian locations. They employed Claude Code to build and test the software, bypassing geographic restrictions through commercial VPNs. Anthropic subsequently banned the associated accounts after identifying the activity as weapons development.
In a separate finding, Anthropic uncovered a state-linked Russian cyberespionage campaign that leveraged Claude to automate infrastructure setup, phishing, and malware development. This operation targeted over 20 organizations, including government, military, and intelligence entities across Ukraine and Europe.
Additionally, the report details the use of Claude in propaganda efforts, specifically a state-directed campaign in the Central African Republic. This initiative generated pro-Russian and pro-Wagner content for local media, radio broadcasts, and Telegram channels.
Anthropic also reported five instances where researchers, potentially linked to military or state-backed programs, used Claude for biological research. The company noted that while it maintains safeguards, current AI models can no longer be guaranteed to remain below the threshold for assisting in sophisticated biological research.
