Chinese Military Researchers Used Anthropic's Claude for Taiwan Air Defense Suppression, Anti-Torpedo Systems, Anthropic Reports

Anthropic's September 2026 threat report details that hundreds of China-linked agents allegedly used the company's Claude model for at least five distinct programs: two military, two surveillance, and one aimed at distilling Claude's capabilities.

One China-based actor used Claude to draft a fire-control specification for an anti-torpedo fire-control system, testing the system against U.S. Navy anti-torpedo and anti-submarine systems based on publicly available information. The actor disguised itself as a U.S. defense OEM, but Anthropic believes the actor was associated with a Chinese defense manufacturer seeking to develop a system for the People's Liberation Army Navy. The actor also prepared a 200+ page technical proposal for a potential client.

Another China-based defense researcher used Claude to develop about 16 software modules for electronic warfare and suppression of enemy air defenses. The software analyzed radars, SAM sites, command posts, and communications nodes, and prioritized targets. Anthropic noted that the default scenario contained 12 targets in Taiwan, including Patriot and Tien Kung batteries, air bases, an early-warning radar, and a command bunker. Account metadata and content caught by safeguards indicated the actor was linked to PRC research institutions, including the PLA Academy of Military Sciences, though Anthropic does not outright state that Claude was used by the PLA.

Anthropic also disrupted China-linked surveillance operations related to Uyghurs outside of China. In one case, a China government-linked actor used Claude to infiltrate Uyghur networks abroad.

The report raises questions about why Chinese actors would use a U.S. frontier model over domestic alternatives. Given Chinese-language prompts and account-level evidence, plausible deniability is unlikely the primary reason. More likely, Claude was better or more convenient for these particular engineering workflows, particularly coding, reasoning, and agentic tasks. Additionally, U.S. frontier models trained on English-language material may have more extensive knowledge of publicly available information about American military technologies and systems.

Why This Matters

The alleged use of a U.S. AI model by Chinese military researchers for Taiwan-specific targeting and anti-torpedo development raises national security concerns about dual-use AI technology.
It highlights the challenge of preventing state-aligned actors from accessing advanced AI models for military purposes, even with safeguards in place.
The report underscores the competitive gap between U.S. and Chinese frontier models in certain engineering and reasoning tasks, potentially influencing export controls and model access policies.

Background

Anthropic is a U.S.-based AI safety company that develops the Claude model family. The company publishes regular threat intelligence reports detailing misuse of its models. U.S. export controls restrict the transfer of advanced AI technology to China, but cloud-based access to frontier models via APIs remains a vector for potential misuse. China has its own advanced AI models, including those from Alibaba and other tech giants, but reports like this suggest they may not match U.S. models in tasks requiring reasoning, coding, or specific English-language knowledge. The People's Liberation Army has been linked to military AI development, and Taiwan scenario planning is a recurring element in PLA exercises.

Key Perspectives

Anthropic: The company actively monitors for misuse and proactively disrupts accounts linked to prohibited military applications. Their safeguards caught the activity, and they reported it as part of their threat intelligence program.

Chinese Government/PLA: China has denied using foreign AI models for military purposes and claims it develops its own AI capabilities. The report does not name specific officials or confirm direct PLA use, and Chinese actors may use such tools for research or competitive analysis rather than operational deployment.

Critics/Skeptics: Some may question why Chinese military researchers would use a U.S. model that logs their activities, rather than secure domestic tools. The report's reliance on account metadata and prompt language leaves room for false flags or spoofing, though Anthropic considers this unlikely.

What to Watch

Whether U.S. policymakers tighten API access to frontier models for entities linked to China or military research.
How Anthropic and other AI providers update their safeguard systems to detect and block state-linked military use.
China's response—likely denial or dismissal—and whether it accelerates domestic model development to close the perceived gap.

ZOTPAPER