The U.S. National Security Agency, Cybersecurity and Infrastructure Security Agency, and Federal Bureau of Investigation accused six Chinese AI firms of waging industrial-scale attacks to distill capabilities from U.S. frontier models.
The firms, including DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI, are said to have extracted capabilities from models like Claude, GPT, Gemini, and Grok since at least late 2024.
The agencies claimed that Chinese firms likely acted with government awareness when extracting capabilities from U.S. models, allowing them to shorten AI development timelines and reduce financial costs. The U.S. warned that these actions threaten its lead in the AI race, calling for coordinated action across the AI ecosystem to combat the alleged theft.
Attack methods include exploiting AI model inference APIs by bulk-buying fake accounts, which execute highly coordinated queries. Another common method is using prompt injection techniques to jailbreak models, including crafting prompts that force models to reveal their internal reasoning. For instance, DeepSeek used prompts instructing models to imagine and articulate the internal reasoning behind completed responses step by step.
To mitigate the threat, U.S. agencies recommended improving detection of sophisticated campaigns using tens of thousands of accounts and flagging suspicious subscription-to-usage ratios. They also suggested altering model responses subtly to reduce the payoff for Chinese firms, such as presenting correct information with different reasoning or reducing reasoning depth.
The agencies acknowledged that these steps could frustrate legitimate users, who might be switched to inferior models without notice. They also noted that Chinese firms use aggressive, adaptive discovery to identify valuable data, which they then collect to generate synthetic training datasets.
The U.S. warned that without cooperation, it would face ongoing financial harm from the systematic extraction of proprietary capabilities.
Source: arstechnica
