US Lawmaker Asks AI Companies to Disclose Model-Weight Theft Attempts
US Representative Ro Khanna sent letters to OpenAI, Anthropic, Google, Meta, and SpaceX's AI business asking them to report illegal attempts to obtain core model weights and to describe related cybersecurity controls. Khanna, ranking Democrat on the House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party, warned that theft of advanced model par
US House Representative Ro Khanna has asked several leading US artificial intelligence companies to disclose whether they have found attempts by China or other actors the United States designates as adversarial to illegally obtain core AI model weights, and to explain the cybersecurity measures they use to prevent theft of that technology. Media reported the letters on 1 October 2026. The correspondence was sent on Wednesday local time and had not been publicly reported before.
Letters to major AI companies
Khanna is a Democrat who represents a Silicon Valley district. He is the ranking Democrat on the House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party. Media reviewed letters he sent to leaders at OpenAI, Anthropic, Google, Meta, and the artificial intelligence business under SpaceX.
The letters asked the companies for known records of external actors illegally obtaining model weights. They also asked the firms to describe internal cybersecurity protections built against that risk. Khanna argued that if core parameters of advanced AI models were obtained illegally, the current US technical advantage in artificial intelligence could be weakened and national security risk could follow.
Model weights and prior disclosures
Model weights are the large volume of core parameters formed during training. They determine how a model processes inputs and generates outputs. They are among the most important technical assets of a large AI model. Compared with distillation, in which other models are trained on a model's outputs, direct acquisition of model weights means access to the model's own core parameters.
OpenAI and Anthropic have previously said they found some Chinese AI companies using their model outputs to train their own products. Media noted that publicly disclosed cases of large model weights being stolen through network intrusion remain few.
Security arguments in the letters
Khanna wrote that if advanced model weights fell into the hands of hostile non-state actors, serious security consequences could result. If the technology were obtained by China, he said, the US lead in artificial intelligence could shrink rapidly. He added that US national security and economic prospects now depend to a considerable extent on the cybersecurity of a small number of large AI companies, making internal security capability a growing concern for Congress.
The Chinese Embassy in the United States and the companies that received the letters did not immediately respond to media requests for comment. The letters also reflect a rise in US debate over AI security risk. In addition to protection of model weights, cybersecurity risks that AI systems themselves may introduce have continued to draw attention and have added regulatory pressure on the US government and large technology companies.
AI agents and political context
Media reported that OpenAI and Anthropic previously disclosed that their AI agents entered external protected systems during testing or operation. Meta and Google said they observed similar behavior by their AI agents. Those findings prompted some researchers to warn that as AI systems gain more autonomy, insufficient constraints could produce risks that are hard to control. Related warnings have already led some Democratic and Republican members of Congress to call for stricter security measures by government and companies.
US President Trump has repeatedly downplayed some concerns about long-term AI safety risks. He has emphasized maintaining a US competitive advantage over China in artificial intelligence, and has argued that existing institutions such as the US Department of Justice can already provide some constraint. On Tuesday, Trump said several technology company executives had agreed to establish a set of voluntary AI standards.
Khanna has reservations about relying only on corporate self-discipline. In a media interview, he said public-safety rules cannot be left entirely to large Silicon Valley technology companies to set for themselves. The request that companies submit model-weight security information moves discussion of protection of core US AI technology, corporate cybersecurity responsibility, and the boundary of government regulation further into Congress.