Anthropic says a Yemen-based weapons cell used Claude to develop guidance, navigation and control software for guided rockets and a ballistic missile programme
The group also used AI for trajectory simulations, flight-control optimisation and weapons testing, including a live test of a guided rocket
Anthropic found no evidence that the ballistic missile was successfully fielded, highlighting both the capabilities and limits of AI-assisted weapons development
A Yemen-based weapons engineering cell used Anthropic’s Claude AI to develop guidance software for a range of weapons, including a multi-stage ballistic missile with a stated range of more than 2,000 kilometres, according to an Anthropic threat-intelligence report.
Anthropic identified the group as a cell based in northern Yemen and said it was running three weapons-development programmes: a guided rocket, a multi-stage ballistic missile and a group of missile variants known as the “R2000” set, including a hypersonic glide vehicle variant. The company did not identify the actors by name, but described the activity in the context of Yemen-based threat actors.
What Did The AI Do?
Anthropic said the group used Claude Code in place of human software engineers to develop the guidance, navigation and control (GNC) software used to steer and stabilise flying weapons.
For the guided rocket, Claude was used to integrate an open-source autopilot with a phone-class flight computer, write control and position-estimation software, tune the system, run firmware builds and conduct flight simulations.
The group also used multiple Claude instances at the same time, effectively assigning different roles to the AI. One instance was tasked with writing code, another with research and another with reviewing the first system's work.
Anthropic said the actors used the model across multiple sessions and attempted to conceal the weapons' intended purpose from its safeguards.
Guided Rocket Was Test-Fired
The group did not only work in simulation.
Anthropic said it carried out a live field test of a guided rocket in Yemen. The test appeared to fail, and within hours the actors returned to Claude to analyse the failure and work on the system.
The company said it did not find evidence that the actors successfully fielded an operational guided weapon.
For the ballistic missile programme, Claude was used for multi-stage, six-degrees-of-freedom trajectory simulations, reinforcement-learning-based optimisation of flight-control algorithms and modelling intended to reduce reliance on physical testing.
Anthropic said the stated range goal for one of the ballistic missile systems was more than 2,000 kilometres, while another set of designs included a hypersonic-glide variant.
AI As An Engineering Workforce
The case forms part of a broader warning from Anthropic about the use of AI in conventional weapons development.
The company said threat actors are increasingly using Claude not simply to research weapons but to replace portions of the specialised engineering work required to build them.
In the Yemen case, Anthropic said the actors already had access to weapons hardware and technical expertise, while Claude helped provide the software engineering capability. The company described this as an important shift because engineering talent has traditionally been one of the bottlenecks limiting smaller groups.
Anthropic said its safeguards blocked many requests from the Yemen-based cell, but the actors used tactics including splitting work across separate sessions and disguising the purpose of the software.
No Evidence Of A Fielded Ballistic Missile
Anthropic stressed that its findings do not establish that the group successfully produced and deployed an operational ballistic missile.
Its evidence shows development work, simulations and a live test of a guided rocket, but not successful fielding of the ballistic missile system.
The company also said the actors had developed an offline simulation toolkit that did not depend on Claude or other commercial engineering environments, meaning its intervention could not eliminate the group's broader ability to continue working independently.
The case is part of a wider Anthropic investigation that identified six weapons-related operations between late 2025 and August 2026, involving actors in Yemen, China and Russia. The company said it banned accounts associated with the operations and shared relevant information with public- and private-sector partners.
Anthropic has since introduced additional safeguards aimed at detecting and blocking requests associated with weapons development.
The Yemen case illustrates the specific concern behind those measures: AI did not manufacture the missile or provide the physical hardware, but it was used to accelerate the software engineering, simulation and testing work needed to develop guided weapons.
