China-Based AI Companies Using Distillation at Scale Against US AI Companies (www.cisa.gov)

China-based AI companies are actively engaging in large-scale knowledge distillation campaigns aimed at extracting proprietary functionalities from U.S. AI models, including those from prominent companies like OpenAI and Anthropic. This malicious activity, identified and warned about by U.S. cybersecurity agencies such as the NSA, CISA, and FBI, poses a significant threat to U.S. technological leadership. Companies like DeepSeek and Alibaba have reportedly extracted billions of tokens from various models, utilizing sophisticated tactics to bypass geographic restrictions and obfuscate their activities. These strategies not only enable unauthorized access to cutting-edge AI technologies but also support the rapid development of competitive models that could disrupt the AI landscape. The implications for the AI/ML community are profound; such systematic extraction undermines fair competition and innovation, as it allows China-based companies to train their models with advanced capabilities without incurring the steep costs typically associated with AI development. To counter these tactics, U.S. AI firms are advised to implement rigorous detection mechanisms for unusual behavior and anomalies, adapt their responses to limit the effectiveness of malicious distillation, and enhance cross-organizational intelligence sharing. The response strategy suggests a need for a collaborative effort among government and private sectors to safeguard proprietary AI technologies against these aggressive distillation tactics.