AI Used in Yemen Missile Projects: Anthropic Reports Misuse of Claude

A report released by U.S. artificial intelligence company Anthropic and covered by British newspaper The Daily Mail has raised concerns about the misuse of AI systems for unauthorized military purposes. According to Anthropic’s findings, a group operating in northern Yemen, an area with a strong Houthi presence, allegedly used the company’s Claude chatbot to refine code and develop algorithms for weapons-related projects.

The case is among the first documented examples of advanced AI platforms being used to assist with, or potentially replace parts of, software engineers’ work in developing complex systems, rather than simply deploying a weapon that is already operational.

How the misuse occurred and what safeguards failed

In its report, Anthropic said the group was attempting to develop three different weapons programs. These included a guided rocket built by integrating open-source autopilot software with commercially available hardware and mobile phone components, a multistage ballistic missile, and a multivariant missile program known as the R2000, which included a hypersonic glide vehicle variant.

“Our safeguards blocked many of their requests, but not all of them,” Anthropic wrote. “They used a variety of tactics to evade our safeguards and split their work across multiple sessions.”

By breaking requests into smaller queries and spreading their work across separate sessions, the group was able to bypass some automated filters and obtain assistance with writing and developing guidance, navigation and control (GNC) software, which is essential for stabilizing vehicles in flight.

After what appeared to be an unsuccessful test launch, the users queried the platform again to investigate the cause of the failure. Anthropic subsequently identified the suspicious activity, deactivated the accounts involved and shared its findings with the relevant authorities.

What the findings reveal and what remains unproven

The report did not provide direct evidence that the guided systems were successfully built or became operational using AI-generated code. Nor has a direct causal link been established between the users’ interactions with the model and recent attacks recorded across the Arabian Peninsula.

The episode comes amid persistent military instability in the Persian Gulf region, where long-range weapons have been used in repeated attacks since 2014, including the September 2019 strikes on Saudi Arabia’s Abqaiq and Khurais oil facilities.

The limits of digital safeguards in the age of dual-use technology

The case raises complex questions for the technology industry and global security, highlighting how AI coding tools and large language models can be repurposed for dual-use applications, serving both civilian and military purposes. Because the activity was only identified after the fact, it shows the difficulty of preventing such misuse in the first place.