---
title: Anthropic Says a Weapons Cell in Northern Yemen Used Claude Code to Develop Rocket Guidance Software
description: Anthropic says a weapons cell in northern Yemen used Claude Code to write guidance software for a rocket it test-fired and to model a ballistic missile.
author: Darie Nani (Editor-in-Chief)
date: 2026-09-11T21:40:59.801Z
updated: 2026-09-11T21:40:59.809Z
canonical: https://www.sovereignmagazine.com/article/anthropic-yemen-weapons-cell-claude-rocket-guidance
image: https://cdn.nanimediahouse.com/yemen-claude-logo-card-ivory-v1.webp
categories: Artificial Intelligence, Science & Tech
content_type: News
region: Middle East & Africa
publication: Sovereign Magazine
about:
  - type: Organization
    name: Anthropic
  - type: Place
    name: Yemen
---

A weapons engineering cell in northern Yemen used [Claude Code](https://www.sovereignmagazine.com/article/anthropic-claude-code-source-code-leak) in place of human software engineers to write the guidance software for a guided rocket, Anthropic says in [its September threat intelligence report](https://www.anthropic.com/threat-intelligence-report-september-2026). The group then test-fired the rocket and came back to Claude within hours to work out why the test had failed.

The group was running three weapons programmes, according to the report. Besides the guided rocket, which used a phone-class flight computer with homing guidance for the final phase of flight, it was working on a multi-stage ballistic missile with a stated range goal above 2,000 km and a set of missile designs called the "R2000", one of which was a hypersonic glide vehicle.

Anthropic does not name the group or say who it works for, and it says it has no evidence the group fielded a working weapon. Most of northern Yemen, including the capital Sanaa, is held by the Houthi movement, which the US redesignated a foreign terrorist organisation in March 2025.

## Flight Control Software for the Guided Rocket

The group used Claude to fit an open-source autopilot onto the rocket's flight computer, according to Anthropic, writing the control and position estimation software, tuning the control settings, running a firmware build pipeline and simulating a flight. It ran several Claude instances at once and gave each one a job, with one writing the code, another doing research and a third reviewing the code the first one produced.

> "We do not have evidence the actors succeeded in fielding an operational device; but they did test-fire a guided rocket."
> — Anthropic threat intelligence report, September 2026

The field test in Yemen appears to have failed, Anthropic says. Within hours the group had brought the rocket's telemetry back to Claude to find out what went wrong.

## Ballistic Missile and Hypersonic Glide Vehicle Simulations

For the missile programmes, the group used Claude for simulation, modelling multi-stage trajectories in six degrees of freedom for medium-range, intermediate-range and hypersonic glide variants, and tuning flight control using reinforcement learning.

It also used Claude to calibrate its simulations against reference implementations. Anthropic says the aim was a digital model of an operational weapons system that would cut the group's need for physical tests.

Anthropic banned the accounts it linked to the group and shared what it found with public and private sector partners. By then, the company says, the group had already used Claude to compile its simulation toolkit into a standalone program that runs offline, without Claude or engineering software such as MATLAB.

## Anthropic's Safeguards Blocked Many of the Group's Requests

Anthropic says its safeguards blocked many of the group's requests, though not all of them. The group hid its goals and the products the software was meant for, and it split the work across many sessions so that no single conversation showed its full intent. Anthropic found the activity during its own investigations into suspected weapons development.

## Six Weapons Cases in Anthropic's Threat Report

Anthropic describes the use of Claude to develop software for conventional weapons as a new category of misuse since its previous threat report in November 2025. It sets out six cases in the September report, three in China, two in Russia and the one in Yemen.

In one, a small freelance team based in Russia used Claude Code to write software for an autonomous kamikaze drone swarm. In another, a China-based researcher whom Anthropic links to the PLA Academy of Military Sciences used Claude to develop electronic warfare targeting software, and partway through the project changed its default scenario to 12 targets in Taiwan.

Anthropic's Frontier Red Team developed new evaluations alongside the report to measure AI capability in conventional weapons development and tactical intelligence targeting. The company says the evaluations show models making consistent progress on simulated tasks in both areas, and it has launched classifiers designed to detect and block traffic related to high-yield explosives and weapons development.

In the same report, Anthropic says [Moonshot AI and DeepSeek sent their customers' requests to Claude](https://www.sovereignmagazine.com/article/kimi-deepseek-claude-prompts-anthropic) without those customers' knowledge.

## FAQ

**Q: Does Yemen have any missiles?**
Yes. The Houthi movement, which holds most of northern Yemen, has fired ballistic missiles at Israel, including the Palestine-2, first launched at central Israel in September 2024. A UN Panel of Experts has reported systematic violations of the Security Council arms embargo that allow the Houthis to acquire weapons. The cell in Anthropic's report was working on a guided rocket, a multi-stage ballistic missile and a set of missile designs called the "R2000", and Anthropic says it has no evidence the group fielded a working weapon.

**Q: What is the range of Houthi missiles?**
The Houthis say the Palestine-2 has a range of about 2,150 km and can reach hypersonic speeds. Israel's military has said it does not believe they have a hypersonic missile, and analysts at the Atlantic Council assess the Palestine-2 as a locally assembled version of an Iranian design. The cell in Anthropic's report was modelling a missile with a stated range goal above 2,000 km, and Anthropic does not say who the cell works for.

**Q: How is AI really being used in war?**
In its September 2026 report, Anthropic describes six cases in which groups used Claude for weapons work: software for a drone swarm in Russia, an electronic warfare targeting suite and an anti-torpedo fire control specification in China, and the Yemen cell's rocket guidance code, along with procurement and intelligence gathering for weapons programmes. Anthropic says its own evaluations show AI models making consistent progress on simulated weapons development tasks, and it has launched classifiers to detect and block traffic related to weapons development.