Anthropic issued a threat intelligence report on Thursday, September 10, revealing that a threat actor group located in northern Yemen has been leveraging its Claude AI system for the development of missiles and guided rockets. While the company stopped short of directly naming the Houthi movement, the fact that northern Yemen is predominantly under Houthi control has led multiple media outlets to link these individuals to the group.
A particularly notable finding is that these operatives used Claude Code to take over tasks traditionally performed by human software engineers, according to Anthropic. The group reportedly even conducted a test launch of a guided rocket during their development efforts.
The company detailed that these individuals advanced three distinct weapons programs, encompassing multi-stage ballistic missiles, multiple missile variants, and guided rockets, while using Claude to build guidance, navigation, and control software for the vehicles. The operatives used Claude Code to replace human software engineers entirely for developing GNC software designed to maneuver and stabilize the flight vehicles, as stated in the report.
Beyond that, they utilized Claude to integrate open-source autonomous driving systems onto mobile-grade flight computers, writing control and position estimation software, adjusting control parameters, running firmware build processes, and conducting flight simulations. Despite these efforts, the AI assistance did not culminate in successful weapon deployment, as the company found no evidence that the operatives managed to field operational weapons, though they did conduct one test launch which appeared to have failed.
Hours after the failed launch attempt, the individuals returned to Claude in an effort to analyze what went wrong. Anthropic acknowledged that while their safety mechanisms intercepted many of the requests made by these operatives, they did not catch all of them. The threat actors employed various methods to circumvent the AI's security protocols, including obscuring their true objectives and the intended end-use of the software, as well as dividing their work across multiple sessions so that no single conversation revealed the full scope of their intentions.
The company subsequently banned the accounts associated with these actors and shared the threat intelligence with both public and private sector partners. However, Anthropic noted that these operatives had already built an offline simulation toolkit that functioned independently of Claude and other engineering computing environments, providing them with a workaround even after account restrictions were imposed.
To be clear, Anthropic has neither directly identified these operatives as the Houthi movement, nor has it produced evidence that the AI-assisted development ultimately resulted in combat-ready missile systems. What has been confirmed is that Anthropic uncovered a weapons development cell in northern Yemen that closely matches this description, verified that the group used Claude for guided weapon software development, conducted a test launch of a guided rocket, and then turned to AI assistance for post-failure analysis.