Bipartisan Bill Aims to Develop Federal Guidance Against AI Distillation Attacks
A new bipartisan House bill, the AI Cyber Defense Act, directs NIST to create voluntary industry standards and best practices to combat AI distillation attacks, particularly from foreign adversaries.
A bipartisan legislative effort in the U.S. House of Representatives seeks to establish federal guidance for defending artificial intelligence models against sophisticated "distillation attacks." The proposed AI Cyber Defense Act, introduced by Representatives Pat Harrigan (R-N.C.) and George Whitesides (D-Calif.), mandates the National Institute of Standards and Technology (NIST) to develop voluntary industry standards and best practices. The primary goal is to prevent foreign entities, specifically from Russia, China, Iran, and North Korea, from illicitly acquiring and replicating advanced U.S. AI technologies.
Representative Harrigan articulated concerns that adversaries are strategically choosing to lag behind U.S. AI advancements, allowing American companies to bear the high costs of research and development. He posits that these nations can then "distill and steal" these innovations for a fraction of the cost, potentially leapfrogging U.S. efforts once market dynamics shift. This approach incentivizes foreign actors to exploit the intensive R&D efforts of U.S. firms, thereby undermining American technological leadership and economic competitiveness in the AI sector.
The bill requires NIST to first clearly define what constitutes an AI distillation attack, distinguishing it from legitimate collaboration and sharing within the open-weight and open-source AI communities. Distillation attacks involve training a new AI model on the outputs and reasoning processes of an existing, more advanced model. This technique allows attackers to replicate the capabilities of powerful proprietary models without undertaking the extensive and costly development process themselves.
U.S. AI companies, including Anthropic, OpenAI, and Google, have previously accused Chinese entities such as Moonshot AI, DeepSeek, and Alibaba of conducting systematic campaigns to distill and copy their commercial models. These alleged campaigns reportedly involve the use of numerous compromised accounts and false identities to circumvent geographic restrictions and access sensitive U.S. AI models. Thousands, if not millions, of prompts are allegedly used to extract valuable data and capabilities for training foreign AI models.
In parallel, U.S. AI developers face ongoing legal challenges from creators, artists, and media organizations who claim their works were used for training commercial models without consent or compensation. The proposed legislation aims to address the external threat of intellectual property theft while the industry grapples with internal ethical and legal debates surrounding data sourcing for AI training.
As part of its mandate, the bill also proposes the establishment of an "AI Cybersecurity Roundtable" under NIST. This roundtable would convene government officials and AI industry stakeholders to collaboratively develop voluntary standards for the secure and defensive application of AI within critical infrastructure sectors. Representative Harrigan envisions this forum as a crucial space for incorporating independent, third-party expertise from frontier AI companies and cybersecurity professionals.
The legislation defines "entities of concern" as foreign individuals or organizations based in or directed by companies from Russia, China, Iran, or North Korea. The NIST director would also have the authority to designate specific entities as such. The bill's proponents believe that by formalizing guidance and fostering industry collaboration, the U.S. can better protect its AI innovations and secure critical infrastructure against evolving threats.