Chinese firms accused of siphoning US AI secrets
US intelligence agencies have accused a specific list of Chinese tech firms, including DeepSeek and Moonshot AI, of executing industrial-scale distillation campaigns to copy American frontier models. Engadget reports that these companies allegedly stole billions of tokens from systems like Claude, GPT, and Gemini to power their own advanced releases, such as Kimi K3 and DeepSeek R1.
According to Engadget, the joint advisory specifically identifies DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI as the entities accused of extracting billions of tokens from US frontier systems since 2024. The report also highlights that Moonshot’s Kimi K3 model was allegedly trained using data extracted from Anthropic’s Claude Fable, while DeepSeek’s R1 relied on capabilities from multiple American competitors.
Source: EngadgetAccording to a joint advisory from CISA, NSA, and FBI cited by GN technics/ai (en-US), the US government has identified specific Chinese entities—including DeepSeek, Alibaba, and Moonshot AI—that have used industrial-scale knowledge distillation to extract reasoning and coding capabilities from frontier models like Claude and GPT since late 2024. The agencies warn that these operations, which likely occur with Chinese government knowledge, utilize API proxies and account pools to bypass security measures, challenging the previously reported low training costs of firms like DeepSeek.
Source: GN technics/ai (en-US)US intelligence agencies warn that Chinese companies are systematically extracting capabilities from American AI models, a practice that threatens technological leadership and national security.
Source: GN technics/ai (en-US)
