Overview
Ongoing topic tracking adversarial use of frontier AI — especially coding agents — for conventional weapons development, simulation, and testing workflows.
Timeline
- 2026-09-13: Anthropic September 2026 threat report discloses gtg-87001 — northern Yemen cell using claude-code for guidance-navigation-control on guided rockets, ballistic missiles (>2,000 km), and R2000 hypersonic glide variants (2026-09-13-anthropic-threat-report-september-2026-primary)
- 2026-08: Activity disrupted; offline toolkit compiled before ban
- 2026 (field test): Guided rocket test in Yemen appears to have failed
Key Players
- anthropic — threat detection, disruption, reporting
- gtg-87001 — adversarial cell (internal Anthropic designation)
- claude-code — primary tool abused
Analysis
The GTG-87001 case marks a qualitative shift: AI coding agents moved from hypothetical weapons questions into full development-and-testing workflows including live field tests and post-failure diagnosis. Safeguard evasion and parallel-agent-workflows demonstrate that account bans alone are insufficient when actors compile offline toolkits. Connects to pace-the-frontier-framework and embedded-evaluators policy debate.
Anthropic did not confirm operational deployed weapons. Describe actors as "Houthi-linked/northern Yemen cell" per primary source wording.