Russia used Claude AI to develop autonomous killer drones for strikes on Ukraine

What is known?

Bypassing restrictions. The operation, called DronDoc or Serafim, began in May 2026. The developers bypassed Anthropic’s geographical restrictions by using commercial private virtual servers.

Shared memory and coordination. The AI helped the hackers create a swarm coordination system and a shared sliding memory. This allows the drones to operate in coordination without an operator.

Autonomous target selection. A small language model was integrated into the drones. It allows the devices to independently decide whether to attack, conduct surveillance, or return to base, as well as issue a command to detonate without human confirmation.

Training on video footage produced by Ukraine. To recognize objects, the developers trained computer vision using footage of combat operations in Ukraine, dividing targets into friendly and enemy.

Notably, the Russians explicitly included people on the list of targets to be destroyed.

Testing locations. During simulations, the developers repeatedly identified locations in the Donetsk region and frontline areas of Ukraine as test sites.

Who is behind the development?

Anthropic believes that the project was carried out by a small team of freelancers linked to a regional university and a research center of the Russian Academy of Sciences, rather than by direct government agencies.

The attackers claimed to have received funding from the Advanced Research Foundation and the Russian Defense Ministry, but the company was unable to independently verify these claims.

All nine identified developer accounts have already been blocked.

The report also notes that Russia-linked groups are increasingly using AI to carry out phishing attacks against Ukrainian military personnel, hack WhatsApp accounts, and infect networks.