Six Chinese AI firms accused of aggressively copying US frontier models
US urges AI firms to ID, then secretly switch, Chinese users to less-capable models.
TL;DR
- US agencies NSA, CISA, and FBI accuse six Chinese AI firms (DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, Z.AI) of industrial-scale attacks to distill capabilities from US frontier AI models.
- These attacks allegedly exploit AI model inference APIs using fake accounts and prompt injection techniques to extract proprietary functionalities.
- Recommended mitigations include improving detection of sophisticated campaigns, monitoring anomalous prompts and accounts, and 'dumbing down' model responses for suspected attackers.
- US agencies acknowledge these mitigations may frustrate legitimate users and pose technical challenges due to Chinese firms' adaptive discovery methods.
- The US urges AI firms and allied governments to share information to track evolving attacks and prevent significant economic losses.
- China denies the accusations, stating its AI success is due to self-reliance and that US claims are "groundless" and a "smear" campaign.