How the Houthis used AI to build missile software

A weapons engineering cell in Houthi-controlled northern Yemen did not need to recruit a full team of software engineers to work on its guided missiles. According to Anthropic, it turned to Claude.

The cell used the artificial intelligence system to develop guidance software for three weapons programs, including a multistage ballistic missile with a stated range goal above 2,000 kilometers and a missile family that included a hypersonic glide vehicle, a maneuverable weapon designed to travel at more than five times the speed of sound. Claude helped write code, conduct research, review technical work, and simulate flight. Anthropic said the actors operated several instances simultaneously, assigning them roles resembling a small engineering team.

The episode, disclosed in Anthropic's September 2026 threat intelligence report, offers an early glimpse of a different military use for generative AI. Much of the debate around AI and warfare has focused on autonomous weapons and battlefield decision-making. But powerful civilian AI systems could also alter how weapons are developed by reducing the cost and scarcity of the technical expertise required to design, test and improve them.

An AI engineering team

Anthropic identified six cases over the past year in which actors in China, Russia and Yemen used Claude to support weapons programs. Four involved weapons software, including guided rockets, an anti-torpedo system and targeting software for electronic warfare.

The Yemen case was particularly extensive. Anthropic said the cell used Claude Code in place of human software engineers to work on guidance, navigation and control software, the systems responsible for steering and stabilizing a vehicle in flight. Anthropic did not identify the cell as Houthi, but the Associated Press reported that it was based in Houthi-controlled northern Yemen.

The actors ran several Claude sessions simultaneously, using one to write code, another for research and a third to review the work. Claude also helped adapt open-source autopilot software to inexpensive computing hardware, estimate a rocket's position and movement, and simulate its flight. The group eventually created a simulation toolkit that could operate offline without Claude or MATLAB, widely used engineering software for modeling and simulation.

Anthropic found no evidence that the group produced an operational weapon. A guided rocket test appears to have failed, and within hours the actors returned to Claude to diagnose the problem.

The failure illustrates AI's limits. Claude could help write and troubleshoot software, but producing a reliable missile still required hardware, engineering experience, testing and physical integration.

Making expertise cheaper

Claude was doing more than retrieving technical information. It was writing code, conducting research and reviewing engineering work, taking on tasks that would ordinarily require trained engineers and technical specialists.

Anthropic's September 2026 research found that frontier models could perform some military and intelligence tasks historically requiring scarce, highly trained specialists. Intelligence targeting, for example, requires expensive human labor to search, connect and interpret large quantities of information. AI could make some of that labor less scarce.

In its 2026 report, How Artificial Intelligence Could Reshape Four Essential Competitions in Future Warfare, RAND, a nonprofit research organization with extensive work on defense and security, considers a future in which AI can perform cognitive tasks once limited to humans, making human expertise less of a constraint on military operations.

The implications are not confined to combat. RAND argues that AI could also assist with engineering software, military platforms and weapons. Skilled labor is often a bottleneck in production and maintenance, while AI could reproduce some of that expertise at a much lower marginal cost than training and retaining additional personnel.

For armed groups and smaller states, commercial electronics, open-source software and dual-use components may be easier to acquire than the engineers capable of combining them into sophisticated systems. AI cannot supply the missing hardware, but it could reduce that human-capital constraint.

The barriers AI cannot remove

Lowering the cost of technical expertise does not make sophisticated weapons easy to build.

RAND cautions that removing some cognitive constraints does not remove the physical ones. Weapons still face limits involving size, weight, power, materials and hardware, while even highly capable AI cannot compensate for information it does not possess. The Yemen rocket's apparent failure demonstrates the distance between producing plausible engineering work and producing a reliable weapon.

A 2025 report from the Stockholm International Peace Research Institute (SIPRI) and the EU Non-Proliferation and Disarmament Consortium makes this distinction through chemical weapons. AI could lower knowledge barriers for non-state actors, it finds, without solving the separate challenge of turning a dangerous substance into an effective weapon.

Missiles face similar constraints, including manufacturing, components, testing, integration and accumulated practical knowledge. AI can ease access to information and technical assistance without removing those physical barriers.

A new proliferation problem

Governments are only beginning to adapt arms-control and AI-governance frameworks to this possibility.

The SIPRI Yearbook 2026 notes that international debate on military AI was long dominated by autonomous weapons capable of selecting and engaging targets. Since 2023, attention has widened to AI used for targeting, planning and intelligence analysis, while governments have also begun considering the security implications of civilian AI systems.

The Yemen case adds weapons engineering to that discussion and exposes a difficult problem for AI companies. Anthropic said its safeguards blocked many of the cell's requests, but not all. The actors concealed the intended purpose of their work and divided tasks among sessions so that no individual conversation revealed the complete project.

Many of the underlying activities, from coding and simulation to research and engineering review, also have legitimate civilian applications. Distinguishing a benign request from one component of a weapons program can therefore require understanding activity across many interactions rather than identifying a single prohibited question.

AI has not eliminated the barriers separating an armed group from an advanced missile program. The more consequential possibility is that it could lower one of them. If engineering knowledge becomes cheaper, faster and easier to access, the technical capabilities once associated with large defense establishments may become available to a much wider range of actors.