Cursor features in a U.S. congressional report on the adoption of Chinese artificial intelligence models: lawmakers say Anysphere, the company behind the coding platform, built its flagship proprietary model on a Moonshot AI foundation. The report calls for restrictions on models developed in China in federal systems and high-risk functions of critical infrastructure; it does not report that Cursor has decided to remove them.
The joint document was published on November 20, 2026 by the House Committee on Homeland Security and the Select Committee on China. It examines how U.S. companies incorporate open-weight models developed in the People’s Republic of China (PRC) into consumer platforms, enterprise software, and systems that access sensitive commercial or personal information.
Anysphere, Airbnb, and DoorDash in the investigation
The committees identified three companies as specific cases: Airbnb, Anysphere, and DoorDash. According to the report, Airbnb uses Alibaba models in customer service; Anysphere developed Cursor’s flagship proprietary model on a Moonshot AI base model; and DoorDash reported that it had authorized and deployed several model families from the PRC.
The Record described uses of these models in customer assistance, coding projects, and inventory tasks. The committees’ report identifies the companies and some providers, but does not specify in its summary which particular model DoorDash uses or provide a list of the Alibaba models deployed by Airbnb.
The investigation, launched in April 2026, included requests for documents, briefings with the companies, reviews of their written responses, and classified sessions. The committees also consulted developers of advanced models, cybersecurity officials, federal agencies, industry representatives, and academics. Both The Record and the lawmakers noted that Anysphere, Airbnb, and DoorDash took part in discussions about their technical, commercial, and security decisions.
The report does not present Cursor’s removal as a decision that has already been made. Its focus is the use of Chinese-origin models by U.S. companies and the measures Congress should adopt. The Record said the companies did not immediately respond to a request for comment.
Cost, customization, and performance drive adoption
Lawmakers describe widespread adoption across different sectors of the U.S. economy, but say it is difficult to track. Among the reasons they highlight are low cost, ease of customization, and performance close to that of leading models. The report argues that the United States lacks an open-weight alternative that combines price and capability competitively for the market.
The Record explains that the numerical parameters of open-weight models—the weights and biases—are available for download. The report, for its part, says that the use of models developed in the PRC is widespread in the U.S. economy and difficult to track.
The committees’ concern is not limited to the origin of the software. The report argues that some of the apparent progress of Chinese models relies on capabilities extracted from U.S. models, and that U.S. data-labeling companies help provide data for training PRC models.
In their assessment, the committees cite federal investigations indicating that major Chinese developers may have used proxy networks and fraudulent accounts to collect outputs from advanced U.S. models and replicate their capabilities. The report links these practices to unauthorized distillation, intellectual property theft, and other forms of illicit extraction.
“We should not wait five years to discover that much of the world, and much of the United States, is running on AI shaped by the CCP, leaving policymakers with few options other than another costly rip-and-replace operation.”
— Chairmen Andrew Garbarino and John Moolenaar, in the report’s foreword
Twelve recommendations and restrictions for sensitive systems
The report brings together 12 public policy recommendations. They include creating a national U.S. open-weight model initiative, making trusted U.S. AI the global default, and establishing standards to determine model provenance and compile an inventory of AI components.
Chairmen Andrew Garbarino and John Moolenaar call for federal restrictions to be extended beyond DeepSeek. They propose excluding models developed in the PRC from federal systems and government-contracted work, removing them from high-risk functions in critical infrastructure, and prohibiting federal cybersecurity grants from being used to purchase them.
Other proposed measures include improving information sharing on threats, systematically evaluating Chinese models, and translating the findings into practical guidance. The committees also recommend strengthening the defenses of the Cybersecurity and Infrastructure Security Agency (CISA), creating an AI threat communications channel with developers, and making CISA the single point of contact for reports of AI-related cyber incidents.
The list adds proposals to hold those who illicitly extract capabilities accountable, examine the sale of U.S. training data to Chinese laboratories, work with allies, and govern both open and closed models while avoiding excessive regulation. The committees say they will continue to oversee the issue.
The DeepSeek precedent and the warning about replacement
The call to expand restrictions builds on a previous measure by the Department of Homeland Security: an order that excluded DeepSeek from federal civilian agency systems. The chairmen want the limit to cover other models developed in China and certain high-risk uses, in addition to federal contracts.
The report warns that removing models once they have been integrated can become a lengthy and costly operation. The Record cites the lawmakers’ comparison with the rapid rollout of Huawei and ZTE technology in the market, which, according to the report, came to threaten U.S. networks. The concern is that an established technological dependency may be difficult to reverse.
The context includes a joint warning published in September 2026 by the NSA, CISA, and the FBI about the national security effects of large-scale distillation operations targeting U.S. AI companies. The report also recalls an Anthropic Mythos demonstration for members of Congress in May, focused on the ability of advanced models to find software vulnerabilities.
In June, the Subcommittee on Cybersecurity and Infrastructure Protection held a hearing on AI’s impact on cybersecurity and critical infrastructure resilience.
The price-capability gap at the center of the debate
Mark Montgomery, a cybersecurity expert at the Foundation for Defense of Democracies, told The Record that the report explains the risks of the United States failing to compete seriously in the low-cost, open-weight model market. In his analysis, China, which is more constrained in the leading-model segment dominated by U.S. companies such as Anthropic and OpenAI, is seeking to expand global dependence on its open models.
Montgomery argued that businesses and utilities may prioritize upfront cost over security. Lawmakers, for their part, sum up the problem as a combination of cheap, competitive models, limited visibility into their use, and questions about how some capabilities were obtained.
The response they propose includes a U.S. alternative, provenance standards, systematic evaluations, and coordination with allies. For Cursor, the specific fact in the report is its connection to a Moonshot AI base model; the measures announced by the committees are recommendations directed at the government and Congress, not a removal order announced by Anysphere.