Another case **GTG-87001: Yemen Guided Weapons Cell — How Much of This Should We Actually Believe?** The report attributes three parallel programs to the same group of people: - A terminal-guidance rocket using a phone-class flight computer - A multi-stage ballistic missile with a stated range goal above 2,000 km - An internal designation called "R2000," including a hypersonic glide vehicle variant Their workflow, as described: using Claude Code in place of human software engineers, with multiple instances running in parallel — one writing code, one doing research, one reviewing — integrating an open-source autopilot onto a phone-class SoC, writing control and state-estimation software, tuning parameters, building a firmware pipeline, running simulations. They reportedly conducted a live field test of a guided rocket. It failed. They came back to Claude within hours to debug the telemetry. The report also concedes: no evidence of a successfully fielded device, but an offline simulation toolkit already exists that doesn't depend on Claude or MATLAB. Anyone with even a passing familiarity with weapons engineering can see the problem. **The scale mismatch is enormous.** A 2,000 km multi-stage ballistic missile and a hypersonic glide vehicle are national industrial programs. We're talking about propellant production, thermal protection systems, reentry aerodynamics, guidance filter design, ground control infrastructure, dedicated test ranges. That is not a gap you close by wiring ArduPilot to a Snapdragon and having a language model tune your PID loops. The short-range guided rocket is already a stretch for this kind of setup. The same group simultaneously running a medium-range ballistic missile program and an HGV variant — in Yemen — strains credibility to the point of breaking. And if a Houthi cell with a Claude subscription can build a hypersonic glide vehicle, what exactly does that say about U.S. missile defense investments over the last twenty years? **The evidence chain is almost entirely conversational.** The "live field test" is documented by the fact that a user came back a few hours later asking why it flew wrong. There is no independent range data, no recovered debris, no radar track, no third-party confirmation. Western media ran with "Houthi missile engineering cell" — but the report itself never names the organization. The leap from "someone asked Claude about reentry aerodynamics" to "active hypersonic glide vehicle program" is doing a lot of unacknowledged work. **The model is being credited for the entire weapons chain based on what it's good at producing.** LLMs are excellent at writing flight control code samples, simulation scripts, and post-failure fault-tree analyses that sound plausible. That is not the same thing as compressing a guided weapons software team. A failed field test followed by "why did it crash" debugging looks a lot more like amateur rocketry than missile program development — but it gets narratively upgraded into a weapons development case study. **The R2000/HGV naming is the biggest red flag.** Real weapons programs rarely lay out a clean product family tree — variant names, glide vehicle upgrade paths, the whole roadmap — in conversations with a commercial AI model. What's far more likely: someone was prompting Claude in the register of a program manager, Claude obligingly generated a tidy system architecture with variant designations, and the threat analyst interpreted the output as documentary evidence of an actual program. The model performed the role it was asked to play. That's not the same as the role being real.