Missile Development Meets Claude: Inside A Yemen Rocket Test

Artificial intelligence (AI) is increasingly being used for tasks far beyond writing, coding and research. Anthropic’s latest threat-intelligence report, Detecting and Countering Misuse of AI: September 2026, details several cases in which its Claude models were used in weapons development and military activity, including a guided-rocket test in Yemen, an autonomous drone programme in Russia and weapons-system development in China.

The report covers six conventional-weapons cases involving actors in Yemen, Russia and China. In Yemen, a weapons cell used Claude Code to develop guidance software for a rocket that was later test-fired. In Russia, Claude was used in the development of an autonomous drone swarm, while Chinese actors used the AI model for projects ranging from anti-torpedo systems to electronic warfare.

Importantly, Anthropic does not say that Claude itself launched or autonomously controlled the rocket. The actors used Claude to develop and refine software that formed part of the weapon’s guidance and control system.

Claude's Role In Yemen’s Weapons Programme

The Yemen-based cell, which Anthropic says was operating in northern Yemen, was working on three weapons programmes: a guided rocket, a multi-stage ballistic missile with a stated range goal of more than 2,000 km, and a family of missile variants known as the “R2000”, including a hypersonic glide vehicle variant.

Claude Code was used in place of human software engineers to develop guidance, navigation and control software. The actors used it to integrate an open-source autopilot with a phone-class flight computer, write control and position-estimation software, tune control settings, build firmware and run flight simulations.

The report also shows how the actors tried to get around AI safeguards. They used multiple Claude instances at the same time, with separate sessions handling coding, research and code review. They also split their work across sessions and concealed the broader purpose of the project from the model. Anthropic says its safeguards blocked many requests, but not all of them.

From Software To A Live Rocket Test

The Yemen case becomes more significant because the work did not remain limited to simulations or computer code.

The weapons cell eventually test-fired a guided rocket in Yemen. Anthropic says the test appeared to fail, after which the actors returned to Claude within hours to analyse the problem. Claude was used for work including flight-control firmware, terminal guidance and post-test telemetry analysis.

The report, however, draws an important line around what the company actually observed. Anthropic says it has no evidence that the actors successfully fielded an operational device, although they did conduct the live rocket test. The cell had also developed an offline simulation toolkit that did not depend on Claude or MATLAB.

This makes the Yemen case different from a simple example of someone asking an AI model about missiles. Claude was incorporated into a real engineering workflow, helping with software development, simulation and analysis connected to a physical weapons programme.

Russia And China Show A Wider Pattern

The other cases in the report show that the use of AI extended beyond missile development.

In Russia, Anthropic identified a group of likely freelance actors working on an autonomous FPV kamikaze drone swarm. Claude Code was used to develop software for swarm coordination, target detection, terminal guidance and other functions. The system was designed to identify targets and issue detonation commands without a human being directly involved in the final engagement.

The actors also trained a computer-vision classifier using scraped Ukrainian combat footage and used a fixed location in Ukraine’s Donetsk Oblast as a demonstration strike point. Anthropic assessed the group as a small specialised freelance team rather than a Russian state entity, and said the weapons it observed remained at an early development stage.

In China, one actor used Claude to develop an anti-torpedo weapons system, including a fire-control specification, a technical proposal of more than 200 pages and parts of the fire-control software. Another China-based actor used Claude to build an electronic-warfare and air-defence suppression suite capable of analysing radar systems, assessing jamming effectiveness and ranking targets. The project included simulations involving systems such as Patriot and THAAD and was later adapted to a scenario involving 12 targets in Taiwan.

What Is Anthropic Doing To Counter The Misuse?

Anthropic says it banned the accounts associated with the weapons-related activity and shared relevant intelligence with authorities and industry partners. The company also used the cases to strengthen its safeguards.

It has introduced new classifiers designed to better identify and block activity involving high-yield explosives and weapons development. But the report also highlights the difficulty of stopping such activity: users can conceal their overall objective, break a complex weapons project into smaller tasks and spread those tasks across different conversations.

The cases do not establish that AI systems are independently developing or deploying weapons. They do show, however, that generative AI is already being incorporated into parts of weapons engineering, simulation, targeting and testing—areas that traditionally required specialised technical expertise.