New Delhi: A weapons engineering cell in northern Yemen used Anthropic’s Claude AI coding tool to develop guidance software for rockets and ballistic missiles, according to a September report by the US-based AI company. The investigation uncovered three weapons development programmes and a failed guided-rocket test, although Anthropic said it had no evidence that the group had successfully deployed an operational weapon.
The findings raise questions about the potential use of commercially available AI tools by armed groups to support weapons development. While the report does not explicitly identify the group as Houthi rebels, their location in northern Yemen points to a possible connection that has not been confirmed by Anthropic.
In its report, Detecting and countering misuse of AI: September 2026, Anthropic described how the actors used Claude Code to develop guidance, navigation and control (GNC) software, which helps flying vehicles maintain stability and follow their intended trajectories.
Three weapons programmes identified
According to the report, the group was developing three types of weapons, including a guided rocket using a smartphone-class flight computer, a multi-stage ballistic missile with a planned range of over 2,000 kilometres, and the R2000 missile series, which included a hypersonic glide vehicle variant.
The group used Claude Code to assist with software engineering tasks, including integrating an open-source autopilot with a phone-class flight computer, developing position-estimation software, adjusting flight-control settings, building firmware and running flight simulations.
Anthropic said the group also operated multiple Claude instances simultaneously, assigning separate roles for coding, research and reviewing software. This allowed the actors to divide the engineering work across different AI sessions.
AI safeguards failed to block all requests
Anthropic said its safeguards blocked many requests but failed to prevent all of the suspected weapons-development activity. The actors allegedly concealed their objectives and the intended applications of the software while distributing their work across multiple sessions to avoid revealing the project’s full purpose.
The company identified the activity through internal investigations into suspected weapons development. It subsequently banned accounts associated with the actors and shared threat intelligence with public and private-sector partners.
However, the investigation also found evidence that the actors had developed an offline simulation toolkit that did not rely on Claude or engineering software such as “Matlab”, potentially allowing some work to continue independently of the AI service.
Guided-rocket test appears to have failed
Anthropic said the actors had test-fired a guided rocket, but the trial appeared to have failed. The group returned to Claude within hours to investigate the failure, according to the report.
The company stressed that it had no evidence the group had successfully fielded an operational device. The findings therefore document an attempt to develop and test guided weapons, rather than confirmed deployment of a functioning missile system.
The report does not establish that Claude was used in any specific attack on Saudi Arabia or that the weapons programmes were operated by the Houthis.