Get Full Government Meeting Transcripts, Videos, & Alerts Forever!
Get email alerts on the Model Security topic
No spam. Unsubscribe anytime.
Experts warn of model-theft, distillation attacks and poisoned datasets
Summary
Witnesses told a House committee that distillation, trade-secret theft, stolen model weights and data-poisoning present acute risks to U.S. AI leadership and national security; they recommended counter-distillation guidance, NSA red teams and whistleblower protections.
Yousef Mahmoud, director of AI and emerging technology at the America First Policy Institute, told the committee that software-side threats are acute and described four high-risk vectors: distillation attacks, trade-secret theft, theft of model weights, and data poisoning that can insert hidden backdoors.
Mahmoud said distillation attacks have evolved to use thousands of small accounts prompting models a small number of times to evade detection and called for intelligence-community monitoring and voluntary counter-distillation guidance led by a Center for AI Standards and Innovation. He proposed authorizing $50 million to $100 million annually for the center, directing the NSA to run red-team exercises across the AI development pipeline, extending DoD lab-security standards, and passing whistleblower protections for AI researchers.
"Right now, Chinese AI developed from stolen American technology is helping Iran target American warfighters," Mahmoud said, framing the threat as both immediate and operational. Committee members asked whether the Bureau of Industry and Security should make certain types of adversarial distillation illegal; Mahmoud said the intelligence community and industry sharing would be central to enforcement and prevention.
The witnesses and members discussed trade-offs between law enforcement, industry best practices, and statutory remedies; none of the proposals were adopted during the hearing, but witnesses urged rapid congressional action.
