Overview
- OSTP Director Michael Kratsios, in a memo Thursday, said foreign actors based mainly in China are running industrial-scale campaigns to extract capabilities from U.S. frontier AI systems.
- The memo describes attackers using tens of thousands of proxy accounts and jailbreaking to flood models with queries, then training cheaper “distilled” copies that learn from the larger model’s outputs.
- The administration plans to share intelligence with U.S. AI companies, help build joint defenses, and explore accountability measures, though it named no specific firms and announced no immediate sanctions.
- The Chinese Embassy rejected the allegations as baseless, and the warning arrives weeks before President Trump’s planned meeting with Xi Jinping in Beijing as officials weigh restricted Nvidia chip sales that have not yet shipped.
- OpenAI and Anthropic previously reported large-scale distillation attempts by China-based labs DeepSeek, Moonshot, and MiniMax, and officials warn that copied models can look strong on select tests yet lose safety guardrails.