Deterring American AI Model Theft Act of 2026

• •Defines 'closed-source AI models' as proprietary models with restricted access, governed by terms of service, and not publicly available.
• •Defines 'model extraction attacks' as unauthorized querying to replicate or train other AI models by circumventing access controls or violating terms of service.
• •Identifies 'countries of concern' as China (including Hong Kong and Macau), Russia, and other nations designated by the Secretary of State.
• •Defines 'entities of concern' as foreign persons or entities headquartered in countries of concern or those conducting unauthorized model extraction.
• •Defines 'fraudulent account network providers' as foreign entities that knowingly provide access to restricted AI models for entities of concern.
• •Excludes from the definition of 'fraudulent account network provider' entities that provide internet access for freedom of expression purposes.
• •Allows the purpose of a model extraction attack to be inferred from query volume, patterns, coordination, and correlation with other AI development timelines.
• •States that authorized model training consistent with terms of service is a legitimate research method and distinct from extraction attacks.
• •Designates the House Committee on Foreign Affairs and the Senate Committee on Banking, Housing, and Urban Affairs as the appropriate congressional committees for oversight.

The bill creates a federal framework to identify and potentially penalize foreign entities that bypass security controls to steal proprietary AI technology, impacting how U.S. AI companies manage access to their models.

Patterns POLISCOPE noticed across the record. These are observations to investigate, not conclusions.