Artificial intelligence is increasingly being used for applications beyond business, coding and research. A new report from AI company Anthropic has revealed that a weapons development group based in northern Yemen used its Claude AI coding tool to develop guidance software for rockets and missiles, including a guided rocket that was subsequently test-fired.
Anthropic said it had no evidence that the group successfully deployed an operational guided weapon. The test-fired rocket appeared to have failed, and the individuals reportedly returned to Claude to investigate the failure.
Three Weapons Development Programmes Identified
In its September 2026 threat intelligence report, Anthropic described a weapons engineering group in northern Yemen that was working on three separate programmes.
These included a guided rocket using a commercially available flight computer, a multi-stage ballistic missile with a stated range target exceeding 2,000 km, and a missile programme involving a hypersonic glide vehicle variant.
The group used Claude Code to assist with guidance, navigation and control software, which helps steer and stabilise a flying vehicle.
The report said the users operated multiple Claude instances simultaneously, assigning different roles to each. One instance handled coding, another conducted research, while a third reviewed the code produced by the first.
This approach allowed the group to use AI for several parts of its development work rather than relying exclusively on human software engineers.
How the Group Bypassed AI Safety Measures
Anthropic acknowledged that its safeguards blocked many requests related to weapons development, but some activities were not stopped.
According to the company, the users attempted to evade restrictions by concealing their objectives and disguising the intended applications of the software. They also divided their activities across multiple sessions, making it harder for any single conversation to reveal the full purpose of their work.
The findings highlight a challenge for AI developers: individual requests may appear legitimate when viewed separately, even when they form part of a larger prohibited activity.
Anthropic said its investigation identified the suspected misuse, after which it banned the associated accounts and shared threat intelligence with public- and private-sector partners.
Guided Rocket Test Appeared to Fail
The report said the group conducted a live test of a guided rocket. However, Anthropic stated that it had no evidence the group had successfully fielded an operational weapon.
The test appeared unsuccessful, with the individuals reportedly returning to Claude within hours to work out what had gone wrong.
Anthropic also found evidence that the group had developed an offline simulation toolkit that did not depend on Claude or other engineering computing environments.
This raises a broader concern for AI safety: removing access to a particular AI service may not completely stop a group that has already developed independent tools and capabilities.
What This Means for AI Safety
The incident underscores the difficulty of preventing advanced AI tools from being misused in sensitive areas such as weapons development.
AI companies rely on usage restrictions, automated detection systems and investigations to identify suspicious activity. However, users may attempt to bypass safeguards by disguising their intentions or spreading related tasks across multiple interactions.
The findings reinforce the need for stronger monitoring, improved safeguards and cooperation between technology companies and security agencies.
Summary
Anthropic reported that a weapons development group in northern Yemen used Claude Code to assist with software for three rocket and missile programmes. A guided rocket was test-fired but appeared to fail. The company said its safeguards blocked many requests, although some misuse went undetected. Anthropic subsequently banned associated accounts and shared threat intelligence, while warning that offline tools could allow development efforts to continue.