In an alert released on Tuesday, the FBI, the National Security Agency (NSA), and the Cybersecurity and Infrastructure Security Agency (CISA) warned that leading Chinese artificial intelligence (AI) companies have been systematically extracting advanced AI models developed by American firms since 2024. These models include Anthropic’s Claude, OpenAI’s ChatGPT, Google’s Gemini, and xAI’s Grok, which the Chinese companies are using to build their own AI systems. A spokesperson for China’s embassy in Washington did not immediately respond to requests for comment. The process of distilling, or extracting knowledge from existing AI models to create new ones, is common in the AI industry. However, it often violates the terms of service of the original model providers. In recent months, American companies and officials have increasingly accused Chinese AI models of being built by distilling American ones. The U.S. has long accused China of stealing intellectual property from American firms, and this report suggests that the distillation efforts may have occurred with the awareness of the Chinese government, though it does not claim direct involvement by Chinese intelligence. The White House has described the U.S. and China as engaged in a race to develop the most advanced AI technologies. President Donald Trump is set to meet with Chinese President Xi Jinping on September 24, and officials from both countries have exchanged accusations about whether China has improperly stolen AI trade secrets. The report specifically named six leading Chinese AI developers: DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI. It details how these companies allegedly distilled specific features, such as “legal specialization optimization” and “coach/assistant capabilities,” from American models. The report emphasizes the scale and complexity of these distillation efforts, suggesting that it is not a minor part of these companies’ development but rather a central component. American companies have made similar allegations in the past, with OpenAI accusing DeepSeek of improperly basing its model on ChatGPT as early as last year. Google also reported being targeted by attempts to clone its Gemini model but did not name any specific suspects. The alert includes recommendations for U.S. companies, such as verifying users of subscription-based AI services and sharing information about suspicious activities. In some cases, it also suggests modifying responses to potentially harmful queries.