Anthropic has disclosed that a cell of threat actors in northern Yemen used its Claude artificial intelligence (AI) system to help develop software for guided rockets and advanced missiles, raising questions about how commercial AI tools could assist weapons development.
The revelations appeared in the company's September 2026 threat intelligence report, which detailed attempts to use Claude for several weapons development programmes. Separately, Saudi Arabia has faced a renewed wave of missile and drone attacks, including deadly strikes on King Khalid International Airport in Riyadh.
However, the report does not establish that the group was definitively linked to the Houthis, nor does it show that Claude-assisted missiles were used in the recent attacks. Anthropic said it had not found evidence that the developers successfully deployed an operational weapon, although one guided rocket was test-fired.
How Anthropic Says Claude Was Used in Yemen's Missile Programmes
In its September threat intelligence report, Anthropic identified the group as GTG-87001 and said it was operating three weapons development programmes. The company located the activity in northern Yemen, an area associated with Houthi control, but did not publicly confirm the group's identity.
One programme involved a guided rocket using a commercially available, smartphone-class flight computer. Another concerned a multi-stage ballistic missile with a stated range target exceeding 2,000 kilometres. The third involved a family of missiles known as the R2000, including a proposed hypersonic glide vehicle variant.
Anthropic said the developers used Claude Code to produce guidance, navigation and control software, which governs a flying vehicle's movement and stability. According to the company, the system assisted with software development, position estimation, control adjustments and firmware builds.
The group also treated multiple AI sessions as a team of virtual specialists. Anthropic said different Claude instances were assigned tasks such as writing code, conducting research and reviewing the work produced by another instance.
The developers reportedly concealed their intentions and divided tasks across separate sessions, making it harder for individual interactions to reveal the full purpose of their work. Anthropic acknowledged that its safeguards blocked many requests, but some prohibited activity continued.
The company said it subsequently banned accounts associated with the activity and shared relevant intelligence with public and private sector partners.
Why the Failed Rocket Test Matters for AI Weapons Risks
After the group tested its guided rocket, Anthropic said the test appeared to have failed, and the developers returned to Claude within hours to investigate the problem.
The company also disclosed that the developers had built an offline simulation toolkit that could function without Claude or engineering platforms such as MATLAB.
That detail matters because banning an account does not necessarily remove knowledge, software or equipment already acquired by its users. Once developers have obtained useful information, they may be able to continue their work using other tools.
The findings also need to be separated from the current conflict. Reuters reported on 12 October that Saudi Arabia was facing further missile and drone fire after a deadly strike on Riyadh airport, which the Houthis claimed.
The airport remained closed as authorities dealt with the consequences of the attacks. The report does not establish that Claude was involved in those strikes.
Anthropic's disclosure is part of a wider set of cases involving attempts to misuse AI. Its report also examines activity involving cyber operations, surveillance, fraud and other forms of weapons development.
For AI developers, the challenge is not simply preventing a model from answering an obviously dangerous question. Users can disguise their objectives, divide complex projects into smaller tasks and combine AI assistance with tools they already possess.
The Yemen case illustrates the challenges of enforcing AI safeguards. Anthropic's findings document an effort to use AI in missile engineering, but they do not establish that the technology produced a successful operational weapon or contributed to the recent attacks on Saudi Arabia.
