Ukraine war briefing: Russian developers used AI to build ‘kamikaze’ attack drone software, Anthropic says
TL;DR
Anthropic's new threat intelligence report describes a small team of Russia-based developers that used Claude Code to write and test software for a swarm of FPV kamikaze drones. The swarm was meant to pick targets and detonate without human approval. For target recognition the team trained a computer vision model on Ukrainian combat footage scraped from the web, sorting equipment into friendly and enemy categories. Anthropic says the work stayed early stage and was mostly validated in simulation.
Nauti's Take
The real progress here sits on the defense side: Anthropic caught the misuse itself, shut it down and published the details. That kind of transparency is what lets the industry spot patterns early.
The risk is still serious, because if a small freelance team can prototype weapons software with coding agents in weeks, the barrier to entry drops for everyone. Teams shipping AI tools should write their own misuse rules now, not after the first incident.
Sources
