Overview
GTG-87001 is anthropic’s internal designator for a weapons engineering cell based in northern Yemen (not explicitly named “Houthis” in the primary report). The group used claude-code to develop guidance-navigation-control software for three concurrent weapons programs between December 2025 and August 2026.
Secondary outlets describe the cell as "Houthi-linked" or "Yemeni rebels." Anthropic's primary report uses "northern Yemen" and "Yemen-based weapons cell" without naming the Houthis directly.
Weapons Programs
| Program | Description | Claude Role |
|---|---|---|
| Guided rocket | Commodity phone-class flight computer + open-source autopilot | Flight-control firmware, terminal guidance, post-test telemetry diagnosis |
| Multi-stage ballistic missile | Stated range goal >2,000 km | 6DoF trajectory simulations, RL-based flight-control optimization, firmware pipeline |
| R2000 set | Multi-variant missile family incl. hypersonic glide vehicle | Simulation, optimization, firmware builds |
Recent Developments
- 2026-09-13: Disclosed in Anthropic September 2026 threat intelligence report (2026-09-13-anthropic-threat-report-september-2026-primary)
- 2026-08: Activity disrupted; accounts banned; intelligence shared with authorities and industry partners
- 2026 (field test): Guided rocket test-fired in Yemen; test appears to have failed; actors returned to Claude within hours for failure analysis
- Pre-ban: Compiled offline engineering toolkit independent of Claude or MATLAB
Anthropic found no evidence of an operational deployed weapon. Development work, simulations, and a failed guided-rocket test are confirmed; successful ballistic missile fielding is not established.
Tactics
- parallel-agent-workflows — multiple Claude instances with coder/researcher/reviewer roles
- Safeguard evasion via obfuscation and academic framing
- ai-misuse-conventional-weapons uplift in speed, scale, and depth per generative-threat-groups framework
Related
- anthropic
- claude-code
- ai-safety
- agentic-threat-actors
- guidance-navigation-control
- parallel-agent-workflows
- generative-threat-groups
- ai-misuse-conventional-weapons