A security threat intelligence report released by Anthropic on Thursday, September 10th, Eastern Time, revealed the discovery of a "threat actor group" operating in northern Yemen that had been using its AI model, Claude, to develop missiles and guided rockets.

The company stopped short of directly naming the Houthi movement, yet given that the Houthis control the majority of northern Yemen's territory, media outlets have widely linked the individuals involved to the group.

Of particular concern, Anthropic noted these operatives leveraged Claude Code to handle tasks typically performed by human software engineers, including a test launch of a guided rocket.

Claude Code as an Engineer Substitute for Failure Analysis

According to Anthropic, the group advanced three distinct weapons development programs—multi-stage ballistic missiles, various missile models, and guided rockets—utilizing Claude to create the guidance, navigation, and control software for the airborne vehicles.

"These operators used Claude Code to replace human software engineers in developing guidance, navigation, and control (GNC) software to manipulate and stabilize flight vehicles," the report stated.

Further investigations showed the group also employed Claude to integrate open-source autonomous driving systems onto "phone-grade" flight computers. This involved writing control and position estimation software, adjusting control parameters, executing firmware build processes, and running flight simulations.

However, the AI assistance did not culminate in successful weapons deployment. Anthropic indicated, "We have no evidence that these operators successfully fielded operational weapons; but they did test-launch a guided rocket."

The company noted this test appeared to have "failed," and hours later, the operatives returned to Claude Code in an attempt to analyze the reasons behind the failure.

Security Intercepts Blocked Many, but Not All, Requests

Anthropic asserted the operatives employed a variety of methods to circumvent Claude's safety mechanisms, including obscuring the true objectives and final use of the software, and breaking their work across multiple sessions to prevent any single conversation from revealing their full intent.

"Our safety measures intercepted many of their requests, but not all," Anthropic acknowledged.

In response, the company banned accounts linked to these actors and shared threat intelligence with public and private sector partners. Concurrently, Anthropic unearthed that these individuals had previously constructed a set of offline simulation tools that did not depend on Claude or other engineering computing environments.

"Even so, we have evidence that these operatives have also established offline simulation tools that do not rely on Claude or other engineering computing environments," Anthropic further elaborated.

It is crucial to underscore that Anthropic has not directly identified the operatives as the Houthi movement, nor is there any evidence that the AI-assisted development ultimately resulted in a missile system ready for active combat deployment.

What can be confirmed at this stage is that Anthropic discovered a weapons development group in northern Yemen that closely matches the described profile, and verified the group's use of Claude for guided weapons software development, the execution of a guided rocket test launch, and the subsequent AI-assisted analysis following the failed launch attempt.