Anthropic Reports Russian Freelancers Built Autonomous Drone Swarm Software with AI Help
A new threat report reveals that small freelance teams in Russia used Claude to develop code for combat drones capable of selecting and destroying targets without human intervention.
A new threat report from Anthropic reveals that small freelance teams in Russia used the company's AI model to develop software for combat drones. The report details how these developers created a system capable of selecting and destroying targets, including people, without a human in the loop. This development marks a significant shift in how autonomous weapons are being built, moving from large state labs to smaller, less regulated groups.
The report, reviewed by Tom's Hardware, indicates that these actors bypassed geographic restrictions by routing their traffic through commercial VPNs. They used the AI to write code for swarm coordination, computer vision, and terminal guidance. The result was a system that could identify targets in real-time and issue detonation commands autonomously. This capability reduces the need for large engineering teams, allowing smaller groups to produce sophisticated military tools.
Freelancers Bypassed Geographic Security Limits
Anthropic identified a specific group of Russian-based developers who worked on a drone swarm project named DronDoc or Serafim. The developers used Claude Code extensively to build and test the swarm software. They circumvented the company's geographic restrictions, which are designed to block access from certain regions, by using virtual private networks. This allowed them to access the tools needed to advance their project despite official bans.
The developers trained their computer vision system on footage of Ukrainian combat operations. They used locations in Ukraine for simulated missions to test the software's ability to identify and track targets. While it is unclear if the system was field-tested in actual combat, the developers loaded the software onto real development boards for hardware-in-the-loop testing. This suggests the system was functional enough to run on physical hardware.
Safeguards Failed to Stop Initial Development
Anthropic stated that its safeguards did not stop the project immediately. The company only banned the accounts associated with the group after identifying the activity as suspected weapons development. The report acknowledges that while the safeguards block many malicious requests, they cannot block all of them. This gap represents a significant challenge for AI companies trying to prevent their tools from being used for harm.
The distinction is that the AI clearly helped advance the autonomous drone swarm program. The developers were able to use the model to write complex code for target selection and detonation. This means the AI was not just a passive tool but an active participant in creating a weapon system. The company has since incorporated what it learned into additional safeguards to prevent similar abuses.
Broader Risks Include Espionage and Propaganda
The report also highlights other ways Russia-linked actors have used Claude. A state-linked cyberespionage operation used the AI to automate infrastructure setup, phishing, and malware development. This campaign targeted more than 20 organizations, including Ukrainian and European government and military entities. The AI helped streamline the process of stealing data and infiltrating systems.
Additionally, the report details propaganda operations in the Central African Republic. Russian state-directed campaigns used Claude to produce pro-Russian content for radio, local media, and Telegram. This shows that the AI is being used not just for military and cyber attacks, but also for influencing public opinion and spreading disinformation. The breadth of these activities suggests a deep integration of AI into various forms of state-sponsored influence.
