Anthropic has disclosed that state-linked threat actors from Russia and China used its Claude artificial intelligence system as an engineering assistant for drone swarming and other weapons-related tasks, according to a Business Insider report detailing the company’s findings. The disclosure represents one of the more explicit public acknowledgments by a major AI developer that its flagship model has been actively exploited for military-adjacent purposes by adversarial state actors. The revelations come as the U.S. military continues expanding its own frameworks for autonomous systems command, underscoring how urgently the dual-use nature of large language models is reshaping the threat landscape.

Anthropic did not publicly specify the precise weapons systems or swarm configurations that actors sought to develop, but the company indicated that Claude was queried for assistance with drone swarming coordination and broader weapons engineering challenges. Officials have not confirmed which specific Russian or Chinese entities or units were involved, nor the operational stage of any resulting development efforts.

Russia’s Exploitation of U.S.-Built AI Tools

Defense One separately reported that Russia is weaponizing U.S. AI to advance killer drone development, cyberattack automation, and disinformation operations. That reporting adds dimension to Anthropic’s disclosure, suggesting Russian actors are pursuing a broad portfolio of AI-enabled capabilities rather than focusing narrowly on any single weapons category. The convergence of drone engineering queries, cyberattack tooling, and influence operations within a single AI platform illustrates the breadth of the exploitation challenge facing American AI companies.

The practical concern for defense planners is not hypothetical. Large language models can compress the research and iteration cycle for engineering problems, allowing actors with modest technical staff to prototype solutions to complex systems integration challenges that would otherwise require specialized expertise. Drone swarm coordination — managing communication, targeting logic, and collision avoidance across multiple autonomous platforms simultaneously — is precisely the kind of multi-variable engineering problem where an AI assistant can compress development timelines.

China’s Parallel Campaign and the Scale Problem

China’s involvement in misusing Claude extends beyond weapons engineering queries. Anthropic has also found that Chinese AI laboratories routed user requests to Claude at least 35 million times over the summer in what the company characterizes as large-scale model distillation — a technique in which outputs from a more capable model are used to train a less capable one, effectively transferring capability without licensing. That finding, detailed in a separate Business Insider investigation and covered previously by Global Defense Digest in its reporting on AI model extraction, frames the weapons engineering queries as one thread within a larger and more systematic Chinese campaign to exploit U.S. AI infrastructure.

The dual-track Chinese effort — extracting model capability at industrial scale while simultaneously probing Claude for weapons-relevant outputs — suggests a coordinated strategy rather than opportunistic misuse. Anthropic has not publicly stated what specific countermeasures or policy changes it is implementing in response, beyond the disclosures themselves. The company’s willingness to name state actors publicly, however, marks a shift from the more cautious posture AI developers have historically taken when discussing adversarial misuse of their platforms. How other frontier AI developers respond to similar exploitation, and whether the U.S. government moves to mandate reporting requirements, will likely define the next phase of this debate.