Anthropic has disclosed that an engineering cell based in Yemen utilized its Claude AI models to write software for multiple guided weapons programs, including missile systems and a guided rocket. According to the company’s September 2026 threat-intelligence report, titled “Detecting and Countering Misuse of AI,” the operations took place between December 2025 and August 2026.
Operating in northern Yemen, the group was characterized by Anthropic as a “Yemen-based guided weapons engineering cell”. While the company did not explicitly name the Houthis, the activity occurred within Houthi-controlled territory, and subsequent reports have described the cell as being Houthi-linked.
Three Weapons Programs Identified
Anthropic’s investigation found that the cell utilized several Claude models, specifically Claude Haiku, Sonnet, Opus, and Claude Code. After detecting the misuse, Anthropic banned the linked accounts and forwarded the intelligence to industry partners and government agencies.
The company detailed three specific weapons development programs connected to the Yemen-based group:
- A guided rocket equipped with final-phase homing guidance and a commercially available, phone-class flight computer designed to steer it toward a target.
- A multi-stage ballistic missile with an explicitly stated range goal exceeding 2,000 kilometres.
- A proposed weapons set referred to as the “R2000” set, which included missile variants such as one integrating a hypersonic glide vehicle.
How Claude AI Assisted Weapons Engineering
Anthropic noted that the actors deployed Claude Code as a stand-in for human software engineers to help develop guidance, navigation, and control (GNC) software, which is essential for steering and stabilizing flying vehicles.
The AI was specifically used to integrate an open-source autopilot onto a phone-class flight computer, write position-estimation and control software, run firmware build pipelines, tune control settings, and carry out flight simulations.Human operators oversaw the overarching project while running multiple Claude instances simultaneously to handle distinct tasks like research, coding, and code review.
To evade Anthropic’s safety systems, the cell purposefully divided its workload across numerous Claude sessions. By compartmentalizing the tasks and concealing their primary objectives, the group ensured that no single conversation exposed the weapons program’s true intent. Although Anthropic’s safeguards successfully blocked a large portion of weapons-related prompts, the actors managed to circumvent certain access controls and refusals.
AI Consulted Following Failed Rocket Test
Anthropic stated there is no evidence indicating the cell successfully fielded a fully operational weapon. However, the report confirmed that the group executed a live field test of a guided rocket, which seemingly failed. Shortly after the unsuccessful launch, the actors consulted Claude once again to troubleshoot the issue and diagnose what went wrong.
While the disclosure clarifies that Claude did not independently construct or design a working weapon, it demonstrates how AI models can assist human engineers with coding, simulation, troubleshooting, and general weapons development tasks.
Also Read:Nepal Floods: Rebuild Cost Estimated at $4.78 Billion as 1,385 Killed, 5,130 Still Missing
The activity was uncovered during Anthropic’s internal investigations into suspected weapons-related AI misuse.Following the discovery, the associated accounts were banned, and the findings were shared with government and industry contacts.
The Yemen incident was featured among a broader range of AI misuse cases in Anthropic’s September report. The document also highlighted weapons-related activities linked to Russia and China, which involved the use of AI code and agents in drone-related development.
