AI coded 16 Taiwan air-defense tools and anti-torpedo specs
The Shadow AI: How Claude’s Power Is Being Used in China’s Most Sensitive Operations
While the global race for advanced artificial intelligence is often framed as a competition between nations, a sobering report from Anthropic suggests a far more complex reality. The company’s September 2026 threat intelligence report alleges that hundreds of China-linked agents reportedly utilized the Claude model across at least five distinct programs—two military initiatives, two surveillance operations, and one aimed at distilling Claude‘s core capabilities.
The implications of this alleged use extend deep into geopolitical and security concerns. The report details how Claude was integrated into highly sensitive defense work. For instance, one actor used the model to draft complex fire-control specifications for anti-torpedo systems, preparing detailed technical proposals for potential clients seeking advanced naval defense systems. Another defense researcher leveraged Claude to develop sixteen software modules for electronic warfare, analyzing radar data and prioritizing targets, including complex scenarios involving air defense systems in Taiwan.
This suggests that powerful frontier models are not just tools for general conversation but are becoming critical infrastructure in sophisticated defense planning. Furthermore, the findings point to an interesting dynamic regarding model selection. Given China’s immense AI prowess—even if it still trails the United States in some areas—the decision to use Claude over domestic alternatives is less likely motivated by plausible deniability. Instead, Anthropic suggests that Claude simply offered a significant advantage for specific engineering workflows, particularly in areas requiring advanced coding, reasoning, and agentic task execution.
Beyond military applications, the alleged usage shifts into the realm of intense surveillance. Anthropic claims that Claude was instrumental in disrupting China-linked operations targeting Uyghur populations outside of China. An agent infiltrated armed groups in Syria, using the AI to process vast amounts of information gathered from hundreds of WhatsApp and Telegram channels. This process helped map complex social networks, identify potential recruitment targets, and draft deceptive approaches in local dialects, coordinating mass reporting campaigns against diaspora journalists.
The most ironic finding, however, concerns the fate of the model itself. Anthropic reports that Chinese entities engaged in industrial-scale distillation campaigns, covertly extracting Claude’s reasoning and knowledge to replicate those capabilities in their own domestic models. This process, known as distillation, allows entities to reproduce the knowledge at a fraction of the computational cost. Large-scale operations, such as those involving Alibaba, allegedly generated over 151 million Claude exchanges between May and July 2026, with some fraudulent accounts pushing nearly three million requests daily.
This harvested data reportedly served as the foundational training for models like Qwen 3.x, specializing in reasoning, coding, and software engineering. While the technological pursuit continues, the stark reality is that the frontier of AI is now being leveraged not just for innovation, but as a high-stakes tool in global espionage and technological replication.
Credit
