House Committees Urge Crackdown on Chinese AI Models Embedded in U.S. Companies
A bipartisan congressional investigation found that Airbnb, DoorDash and the SpaceX-owned coding startup Cursor have all built products on Chinese-made open-weight AI systems, and recommended barring the models from federal networks and critical infrastructure.

Two House committees on Friday released a joint investigative report concluding that a widening slice of corporate America has quietly built products on artificial intelligence software developed in China, and recommended that federal agencies strip the technology from government networks and the systems that run the nation's most sensitive infrastructure.
The report, issued jointly by the House Committee on Homeland Security and the House Select Committee on the Chinese Communist Party, caps a seven-month investigation that began in April with letters to Airbnb and Anysphere, the parent company of the AI coding tool Cursor. It names three companies that told the committees they use so-called open-weight models built by Chinese developers: Airbnb, DoorDash and Anysphere. Open-weight models are AI systems whose underlying parameters are published for anyone to download, modify and run on their own servers, as opposed to "closed" systems accessed only through a company's paid API.
Cheap, Capable and Hard to Track
The roughly 50-page document, reviewed by CNBC ahead of its Friday release, lays out five central findings. Chief among them: Chinese open-weight models are now widely embedded across the U.S. economy, but that adoption is difficult for regulators or even the companies' own security teams to track once a model has been downloaded and customized in-house. The committees found cost, ease of customization and near-frontier performance are the main drivers, and concluded that the United States currently has no competitive open-weight alternative at the price and performance the market demands.
The report puts some numbers behind that adoption. Airbnb told investigators that a customer-service assistant built on a customized version of Alibaba's Qwen model handles roughly 80,000 guest and host interactions a day. DoorDash reported that it had authorized 25 separate Chinese model families for use, and defined a "non-trivial" deployment as one that had incurred at least $5,000 in computing and serving costs. Anysphere's flagship Composer 2 model, which powers Cursor, was built on top of Kimi K2.5, an open-weight model released by the Beijing-based startup Moonshot AI.
The committees also tied Chinese advances to what they described as large-scale "distillation," in which Chinese labs train newer models using outputs harvested from American frontier systems through what the report calls proxy networks and fraudulent accounts. That claim echoes a joint advisory the NSA, the FBI and the Cybersecurity and Infrastructure Security Agency issued in September warning of a similar pattern. The report goes further, alleging that personnel tied to a People's Liberation Army cyber unit visited the Beijing headquarters of the Chinese AI firm Zhipu on multiple occasions, most recently in August, and that PLA procurement documents call for integrating DeepSeek's models into a military cyber range designed to support "intelligent attack" and "intelligent penetration" exercises.
Three Companies, One Warning
Of the three companies named, Cursor's situation drew the most attention because of its size and its new ownership. SpaceX completed its acquisition of Anysphere in August at an implied equity value of $60 billion. The report states that Cursor "informed the Committees" it plans to explore phasing out PRC-developed open-weight models because of geopolitical risk, language that falls short of a firm commitment to drop them. Airbnb and DoorDash have also engaged with the investigation, the committees said, but neither company has announced any change to its model usage.
None of the three companies responded to requests for comment from reporters covering the report's release Friday; Airbnb said only that it declined to comment. The Commerce Department's Bureau of Industry and Security and China's foreign ministry also did not immediately respond to outside inquiries.
The report frames the dynamic as a repeat of an earlier technology fight. It compares the spread of Chinese open-weight AI to the earlier spread of Huawei and ZTE telecommunications equipment through American networks, warning that removing Chinese models from U.S. systems years from now, once they are deeply embedded, could take years and cost taxpayers billions of dollars. "Beijing is using stolen intelligence to close the gap," the committees wrote, adding that the risks posed by the current trajectory "can no longer afford to overlook or downplay."
"Congress has too often been behind the curve on emerging technology. The time to act is now, while we still have the opportunity and leverage to do so."
That line, from the report's foreword, is attributed jointly to Homeland Security Committee Chairman Andrew Garbarino, R-N.Y., and China Select Committee Chairman John Moolenaar, R-Mich., who co-authored the document.
Twelve Recommendations, No Legislation Yet
The report sets out twelve policy recommendations rather than draft legislative text. The most consequential would bar China-developed AI models from federal information systems and from use in performing federal contracts, extending a narrower order made public in October that barred the Chinese model DeepSeek specifically from federal civilian agency systems. Other recommendations call for a "national American open-weight AI initiative" that would combine public computing resources, shared evaluation infrastructure and lawful access to American frontier models for training smaller open systems; a model-provenance standard and "AI bill of materials" that would require companies to document which base models, training data and hosting arrangements sit beneath any AI product they sell or operate; and a single federal reporting channel, run through CISA, for AI-related cybersecurity incidents.
The committees also recommended sanctions or other accountability measures against companies and individuals found to have illicitly distilled American models, and called for a systematic, recurring government evaluation program for Chinese models already in wide use, such as DeepSeek, Kimi, Qwen and Zhipu's GLM. Existing American open-weight alternatives cited favorably in the report include Meta's Llama and Muse, OpenAI's gpt-oss, Nvidia's Nemotron and Google's Gemma, though the report concludes none yet matches the combination of price and performance driving companies toward Chinese alternatives.
Outside analysts who have followed the investigation said the report's core warning is less about any single company's choices than about a structural cost gap. Chinese models such as DeepSeek and Qwen have been estimated to run at roughly a sixth to a quarter of the cost of comparable American systems, a gap that research groups focused on China's technology strategy argue Beijing is actively engineering by subsidizing development and giving models away for free. Critics of a heavy-handed response have countered that open-weight models, once released publicly, are difficult to restrict without running into First Amendment protections on publicly available software and research.
No bill has yet been introduced to implement the report's recommendations, and the committees did not set a timetable for follow-up hearings. The findings nonetheless add pressure on the Commerce Department and the White House, which have spent much of the year weighing, and in some cases walking back, broader restrictions on open-weight AI models developed abroad. For the three companies named in the report, and for the wider universe of U.S. firms that have adopted Chinese open-weight systems to cut costs, the immediate practical question is whether Friday's warning becomes the basis for new procurement rules, mandatory disclosure requirements, or simply another entry in a long-running congressional paper trail on the subject.

Anthropic mathematician says he and an AI model may have solved a 167-year-old prime-number puzzle
OpenRouter Pulls 125 Endpoints After Research Finds AI 'Privacy' Promises Were Broken

Thinking Machines releases a laptop-ready AI model built on reused layers
