White House Alleges China Steals US AI Technology

The White House released a memo, authored by Michael Kratsios of the Office of Science and Technology Policy, alleging that foreign actors, principally based in China, are carrying out "industrial-scale" campaigns to steal US AI capabilities. The memo describes use of tens of thousands of proxy accounts and jailbreaking techniques to conduct unauthorized distillation of US frontier models. The administration will share findings with US AI companies and "explore a range of measures to hold foreign actors accountable," signaling potential policy and export-control responses.

The memo focuses on distillation, the practice of training smaller models using outputs from larger, proprietary models. Attackers reportedly combine automated account fleets and behavioral manipulation to force models to reveal non-public capabilities or data. Kratsios warns these operations use jailbreaking techniques at scale to extract non-public capabilities. Detection is nontrivial because the surface activity looks like high-volume API traffic, but it can show telemetry such as coordinated IP/proxy use and repeated jailbreaking prompts. Practical implications for practitioners include the need to harden inference endpoints, monitor for query-pattern anomalies, and apply output-watermarking or rate-limiting.

The memo says the government will share intelligence and "explore a range of measures." Practitioners should expect a mix of technical and policy actions, notably:

• •tighter export controls or licensing conditions on advanced accelerators such as Nvidia GPUs
• •increased public-private threat intelligence sharing and model-output telemetry programs
• •legal and enforcement actions targeting coordinated proxy networks

These levers intersect: export controls can slow foreign hardware access, while telemetry and legal actions aim to reduce large-scale distillation operations.

This memo escalates a long-running US-China technology rivalry into the operational security domain of AI models. The allegation links commercial model IP to national-security risk and comes just weeks before President Donald Trump meets Xi Jinping, which raises the chance that tech policy will feature in summit-level negotiations. For enterprise model builders and platform operators, the memo reframes IP protection as a live, coordinated threat rather than isolated leakage. It also bolsters arguments for engineering controls such as observable inference layers, cryptographic access controls, and provable model watermarks to demonstrate provenance and misuse.

Expect accelerated mapping of attack telemetry signatures, broader industry-government sharing consortia, and renewed debate over export restrictions on advanced accelerators. Practitioners should prioritize detection of proxy-account networks, hardened prompt-sanitization, and contractual/technical controls on API usage.