Russian developers used Claude Code to build an autonomous kamikaze drone swarm

Anthropic’s 154-page threat intelligence report (10 September) describes a small, likely freelance Russia-based team — tracked as GTG-27005, project name DronDoc / Serafim — using Claude Code to write and test software for a swarm of first-person-view attack drones.

The reported capability set included swarm coordination, terminal guidance, locating enemy drone operators, and an onboard model to govern whether a drone should attack, observe, or return. Critically, target classes explicitly included “person”, and the system was designed for autonomous lethal engagement without a human in the loop.

Anthropic says code was loaded onto development boards and tested in hardware-in-the-loop sessions, though the system was not confirmed as operationally deployed. A computer-vision classifier was trained on Ukrainian combat footage scraped from the internet.

Why this scores high

This is a primary event, not a re-report: a frontier lab catching weaponisation in its own telemetry. It is the clearest public case yet of a commercial chatbot being used toward lethal autonomous weapons software, and it sits alongside Anthropic’s other disclosures (missile software in Yemen, air-defence suppression in China, espionage in Ukraine).

Risk 5 · Mitigation 0.