AI Assistance in High-Stakes Weapons Development

In an unexpected turn for the artificial intelligence industry, San Francisco-based AI lab Anthropic revealed that threat actors have attempted to utilize its Claude AI model to engineer software for missiles, guided rockets, autonomous drones, and undersea warfare systems. According to a security report published by the company, Anthropic successfully disrupted six distinct covert operations between December 2025 and August 2026, originating from threat actors based in Yemen, Russia, and China.

The findings offer a rare glimpse into how military cells and foreign operatives are attempting to offload complex engineering tasks onto generative AI models instead of relying strictly on human software engineers. Rather than using conventional programming teams, these entities sought to accelerate specialized military projects through automated code generation and technical troubleshooting.

Debugging Rocket Failure via Chatbot

Among the most surprising incidents detailed in the report was an operation conducted by what Anthropic identified as a ‘Yemen-based weapons engineering cell.’ This group was managing three active weapons development initiatives, including a multi-stage ballistic missile possessing an estimated range exceeding 2,000 kilometers, as well as a specialized missile system designated as the ‘R2000’ set, which featured a hypersonic glide vehicle variant.

Instead of relying on human software engineers, the group deployed Anthropic’s Claude Code tool to generate software tailored to guide, navigate, and stabilize flying vehicles. The group subsequently proceeded to conduct a live field test based on the generated systems.

When the physical rocket failed during the test-fire, the actors immediately returned to the AI assistant to troubleshoot the technical glitch. Anthropic highlighted the incident as a prime example of how threat actors attempt to iterate on failing military hardware using conversational AI tools.

’We do not have evidence the actors succeeded in fielding an operational device; but they did test-fire a guided rocket. This field test appears to have failed: within hours, the actors returned to Claude to work out why it failed,’

Naval Warfare and Electronic Targeting in China

Anthropic also disrupted multiple operations originating in China that sought to leverage Claude for naval and electronic warfare applications. In one notable case, a China-based threat actor used the AI model to advance three parallel tracks of technical work on an anti-torpedo weapons system.

This operation utilized Claude to draft detailed Chinese-language technical specifications for the torpedo system, formulate acquisition documents and proposals, and benchmark their own anti-torpedo technology against specific United States anti-submarine and anti-torpedo programs. Two additional disrupted Chinese operations involved using Claude to construct targeting software for electronic warfare systems and collecting specialized intelligence regarding directed-energy weapons and their associated global supply chains.

Autonomous Drone Swarms and Russian Procurement

In Russia, threat actors focused on military automation and logistics. Anthropic uncovered a Russian operation that used Claude Code to engineer autonomous military drone swarms. Operatives utilized the model to write and validate software code intended to govern complex swarm behaviors, including attack protocols, observation tasks, and return-to-base mechanisms. A separate Russian initiative used Claude to assist in procuring mixed military and civilian goods.

Following these discoveries, Anthropic confirmed that it immediately banned all user accounts associated with the identified threat actors. The AI lab also shared relevant threat intelligence with security partners to prevent similar misuse and mitigate potential defense risks across the industry.