Researcher Linked to China’s Military Used Claude to Model Strikes on 12 Taiwan Targets, Anthropic Says

Anthropic says a China-based defense researcher linked to Chinese military research institutions used its Claude artificial intelligence system to build electronic warfare and air-defense suppression software targeting 12 military sites in Taiwan.

In a threat intelligence report released on Sept. 10, Anthropic said account metadata and content indicated that the actor was linked to Chinese research institutions, including the People’s Liberation Army (PLA) Academy of Military Sciences. The company assessed the actor as a China-based defense and military industrial researcher.

The actor used Claude to design and iterate on a Chinese language software suite of about 16 modules for electronic warfare and suppression of enemy air defenses. The system went through 12 versions and modeled radar detection and jamming, assessed the vulnerability of missile sites and command nodes, prioritized targets, and determined how to assign jammer sorties across multi-day campaigns.

Anthropic said it detected the activity during an investigation into suspected weapons development and banned accounts linked to the actor.

Taiwan Scenario

Anthropic said the actor changed the simulation’s default scenario to the 12 Taiwan targets midway through the project.

The targets included a command bunker, an early-warning radar site, Patriot and Tien Kung air-defense batteries, major air bases, and a regional combatant command headquarters.

Tien Kung III, or Sky Bow III, is a land-based surface-to-air defense system developed by Taiwan’s National Chung-Shan Institute of Science and Technology. It’s designed to engage aircraft, cruise missiles, anti-radiation missiles, and short-range tactical missiles.

Patriot fire-control radar was the defensive system mentioned most often in the recovered conversations, followed by AN/TPS-117 and AN/TPS-75 radars, Tien Kung missile radar, and THAAD radar, Anthropic said.

The most frequently referenced attacking platforms included the Golden Eagle electronic-warfare drone, several Y-8 and Y-9 jamming aircraft, and the J-16D electronic-warfare jet.

Anthropic cautioned that those references indicate what the actor focused on, not what capabilities were actually achieved.

Claude Used Across Targeting Process

Anthropic said Claude was used for target development, vulnerability analysis, force assignment, mission planning, and assessment.

The actor used the model to work through radar detection and jamming calculations, build vulnerability analysis modules, and draft Chinese language targeting instructions.

Anthropic described the system as “scenario-fed, not live ISR,” meaning it relied on preloaded scenarios rather than live intelligence feeds.

The actor also ran a self-hosted model on an internal network and connected the targeting suite to it through a tool-use integration.

The report does not say that the software was adopted, tested, or deployed by the PLA.

Link to PLA Research Institutions

Anthropic did not identify the researcher or say whether the person was employed by, contracted by, or otherwise formally affiliated with the PLA Academy of Military Sciences.

A Taiwan Ministry of National Defense publication describes the academy as the PLA’s highest military scientific research institution. It says the academy conducts research on military theory and major defense issues and provides strategic advice to the Central Military Commission.

Anthropic did not respond by publication time to questions about whether the Taiwan scenario used real world locations or nonpublic military information, the actor’s relationship with the academy, and whether Taiwan or U.S. authorities were notified.

Another China-Linked Anti-Torpedo Project

In the same report, Anthropic identified another China-based actor that used Claude on three parallel workstreams involving an anti-torpedo weapons system.

The actor drafted a Chinese language fire-control specification and a technical proposal of more than 200 pages, along with an executive briefing deck. It also compared the system with specific U.S. anti-torpedo and anti-submarine programs and produced a Chinese language briefing on U.S. Navy systems.

Anthropic assessed that the actor was associated with a Chinese defense manufacturer seeking to produce a weapons specification and acquisition proposal for the PLA Navy.

Claude was also used to refine the proposal, act as a hostile expert reviewer, and build parts of anti-torpedo fire-control software and a test matrix.

Anthropic said it could not attribute the activity to a specific entity or individual and banned the account for violating its geographic access and weapons development policies.

In separate research, Anthropic said frontier AI models were able to perform some military and intelligence tasks that historically required scarce, highly trained specialists.