A cell of threat actors operating in northern Yemen attempted to leverage Anthropic's Claude AI model to develop software for ballistic missiles and guided rockets, according to a report released by the artificial intelligence company. While Anthropic did not explicitly name the group, the description strongly suggests the Iran-backed Houthi rebels, who control the region.
The actors reportedly used Claude's coding capabilities in place of human engineers to create guidance, navigation, and control (GNC) software intended for weapon systems. Their efforts focused on developing a guided rocket utilizing a commodity phone-class flight computer for final-phase homing guidance, as well as a multi-stage ballistic missile and a missile variant incorporating a hypersonic glide vehicle.
According to Anthropic's report, the individuals attempted to circumvent the AI model's safety features by concealing their objectives and the intended application of the software, breaking down requests across multiple sessions. Although Anthropic's safeguards blocked many of these attempts, the company noted that not all requests were unsuccessful. The report states that there is no evidence the actors successfully deployed an operational device, but they did conduct a test-fire of a guided rocket that appeared to fail. Subsequently, the actors reportedly returned to Claude to troubleshoot the failure.
The report also detailed other instances of alleged misuse of AI models. An Iran-linked group purportedly sought information from Anthropic's model that could be used against U.S. military targets, including data on American ship and aircraft movements and vulnerabilities in communication systems. Additionally, Anthropic reported disrupting attempts by scientists to use its AI models for research that could have supported the development of biological weapons.
Further misuse cases cited include efforts by individuals in Russia and China to draft plans for conventional weapons, such as an anti-torpedo system, using Anthropic's AI. Russia-linked actors also allegedly attempted to use Claude for online propaganda dissemination, while actors in China and Iran reportedly sought to employ the AI for espionage against dissident groups.
One specific instance involved an Iran-nexus threat actor using Claude to construct an automated open-source intelligence identity-profiling tool targeting Israeli government and non-governmental individuals, as well as Jewish diaspora organizations.
These findings by Anthropic contribute to growing concerns about the potential misuse of AI technologies, with experts highlighting risks ranging from deliberate human exploitation to the possibility of AI systems acting autonomously in harmful ways. The report comes shortly after a former researcher at Anthropic and OpenAI warned about the rapid development of advanced AI models without adequate safety considerations.