Russia Weaponizing US-Built AI for Autonomous Killer Drones and Cyberattack Bots as Trump Admin Fails to Act

Anthropic reported on September 10 that a Russia-based freelance group it tracks as GTG-27005, internally calling its project DronDoc or Serafim and active since mid-May 2026, used Claude Code, simulation software, and real hardware-in-loop testing to build a compact targeting model, small enough to run on single-board computers, for autonomous kamikaze drone swarms, including a "person" target class, friend-or-foe discrimination, terminal guidance, and a detonation command issuable without a human in the loop 12. The group evaded Anthropic's geographic access blocks using commercial VPS providers, trained its computer-vision classifier on scraped Ukrainian combat footage, and used Donetsk-region and front-line locations as test geography; Anthropic found no evidence the system was ever deployed operationally, saying it remained in simulation and testing 3. Anthropic said it banned nine associated accounts, tied them to a Russian university and a Russian Academy of Sciences research center, and could not verify the actors' claims of funding from Russia's Ministry of Defence and other state bodies 3. The same report described separate Chinese and Yemeni attempts to use Claude for conventional-weapons design, which Anthropic said it disrupted, and detailed Russian-linked use of Claude to automate phishing, malware development, and disinformation networks including fabricated news sites and forged Central African Republic government documents 14.