Yemeni weapons manufacturers used Claude Code instead of a full-time programmer, creating a virtual engineering team to build missile guidance systems. After a failed test launch of a guided missile, the militants returned to the AI model within hours to analyze the telemetry and determine the cause of the failure.
Yemeni weapons manufacturers used Claude Code instead of a full-time programmer, creating a virtual engineering team to build missile guidance systems. After a failed test launch of a guided missile, the militants returned to the AI model within hours to analyze the telemetry and determine the cause of the failure.
This is according to a new security report published by Anthropic. Cybersecurity experts have exposed and completely dismantled six large-scale operations using artificial intelligence to develop conventional weapons and logistics, which were behind the attackers from Yemen, Russia and China.
Among these cases, the case of a paramilitary group from northern Yemen, codenamed GTG-87001, which developed three weapons programs at once: a guided missile with an on-board computer based on a regular smartphone, a multi-stage ballistic missile with a range of over 2,000 km, and the R2000 missile complex with a hypersonic glider stands out. To develop the flight control, navigation, and stabilization (GNC) software, the weapons developers deployed several parallel Claude sessions, distributing roles as in a real IT department.
One instance was responsible for writing the firmware and control code, a second performed research tasks and selected algorithms, and a third conducted a thorough review of the generated solutions. To bypass Anthropic’s defenses, the militants concealed the military mission and broke the task into smaller fragments between separate sessions.
Although the attackers did not create a production model, they conducted a real test launch of a guided missile. After its failure, the developers returned to Claude within a few hours to analyze telemetry errors. In addition, the Yemeni gunsmiths managed to create an autonomous simulation toolkit, compiled into a file, which can run offline without access to Claude or specialized software such as MATLAB.
In addition to the Yemeni cell, the report revealed three Chinese operations aimed directly at developing weapons and firmware for them:
Swarm of combat drones: Chinese developers tested swarm algorithms in simulation, and the code generated by the neural network was downloaded directly to real electronic boards.
Torpedo interception: creation of design documentation, proposals, and software for a naval torpedo interception system.
Air defense suppression: development of specialized software for guidance and targeting in electronic warfare and neutralization of air defense systems.
Two more uncovered operations were connected to Russia and concerned not the creation of the weapon code, but its intelligence and material support. In one case, a neural network was used to organize the procurement of dual-use goods for the Russian defense-industrial complex. In another, the attackers collected information from open sources about directed energy weapons (laser and microwave systems) and their suppliers to form a database of potential targets and technologies.
Anthropic has blocked all identified accounts and shared the data with partners and government agencies. To counter such threats, the company has launched new classifiers that automatically recognize and block masked requests for explosives, targeting algorithms, and military equipment.
In that report, Anthropic also said that over the past eight months it has uncovered several instances of likely malicious use of its Claude models. These include a cyberespionage operation linked to Russia and attempts by Chinese AI companies to extract and copy Claude’s capabilities.
