Technology
U.S. accuses six Chinese AI companies of systematically stealing U.S. model technology
Federal agencies say DeepSeek, Moonshot AI, and four others have conducted large-scale theft of proprietary AI capabilities since late 2024.
Ars Technica
The U.S. government has publicly identified six Chinese artificial intelligence firms engaged in what it characterizes as coordinated, industrial-scale theft of proprietary technology from American AI developers. The companies—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—have allegedly been targeting frontier AI models since at least the final months of 2024, according to a joint advisory released Tuesday by the National Security Agency, Cybersecurity and Infrastructure Security Agency, and Federal Bureau of Investigation.
How the theft allegedly works#
The accused firms are said to employ multiple methods to extract capabilities from U.S.-developed systems, including popular models like Claude, GPT, Gemini, and Grok. One technique involves purchasing large numbers of fraudulent accounts and then bombarding AI model interfaces with thousands or even millions of carefully coordinated queries designed to systematically map out the models' underlying capabilities. These queries often feature identical or near-identical prompts focused on similar topics.
A second approach leverages prompt injection—a technique where attackers craft specially designed inputs to make AI models bypass their safety guardrails. According to the agencies, some firms have asked models to externalize their internal reasoning processes step by step, effectively reverse-engineering how the systems arrive at their answers. The advisory specifically highlighted DeepSeek as having used this tactic.
The accused firms allegedly route these requests through proxy networks spread across multiple geographical locations, making detection and blocking significantly harder for model developers.
Strategic advantage and government involvement#
U.S. officials argue that this wholesale duplication of model capabilities translates into enormous competitive and financial gains. By obtaining the functional properties of frontier models without having to invest in the years of research and billions of dollars required to develop them independently, Chinese firms can dramatically accelerate their own AI development cycles.
The agencies further suggested that these activities likely occurred "with Chinese government awareness," implying state-level coordination or at minimum tacit approval of the theft.
