Key facts
- The House Homeland Security Committee and the Select Committee on the Chinese Communist Party issued a joint 50-page report.
- Report co-authors are Representatives Andrew Garbarino and John Moolenaar, both Republicans from New York.
- Lawmakers received briefings from DoorDash, Airbnb, and Anysphere about model integration.
- None of Airbnb, DoorDash, or Anysphere immediately responded to a comment request.
House lawmakers issued a joint 50-page report on Nov. 20, 2026, cautioning U.S. firms against embedding China-developed open-weight artificial intelligence models — models that publish their numerical weights and biases for anyone to download — in their technology stacks.
The document came from the House Homeland Security Committee and the Select Committee on the Chinese Communist Party. The investigation began in April 2026 and drew on hearings, classified briefings, and agency advisories. Report co-authors are Representatives Andrew Garbarino and John Moolenaar, both Republicans from New York. Airbnb and Anysphere received letters from House committees on April 29, while DoorDash received one on July 31, and the DoorDash letter requested information by August 14.
The report names Airbnb, DoorDash, and Anysphere, owner of Cursor, as users of Chinese open-weight models. Lawmakers received briefings from DoorDash, Airbnb, and Anysphere about model integration. Those companies used the models for customer assistance, coding projects, and inventory. None of Airbnb, DoorDash, or Anysphere immediately responded to a comment request.
Airbnb’s chief executive said, per the South China Morning Post, that Airbnb relies heavily on Alibaba’s Qwen models for its customer service agent. In a May 20 Bloomberg interview, the chief executive said Airbnb does not buy from Alibaba or provide data to Chinese companies. Airbnb told reporters it mostly uses U.S.-origin models plus a limited number of China-origin open-source models through vetted U.S.-based providers.
Cursor built its Composer 2 coding model using Kimi K2.5 from Moonshot AI, and CNBC reported Cursor plans to stop using Chinese models. Artificial Analysis scored Moonshot’s Kimi K3 at 57 points. K3’s hallucination rate climbed to 51% from 39% for K2.6.
New businesses and Fortune 500 firms use Chinese open-weight models because the price is low, the models can be adapted, and they perform well. Many organizations that use these models say low cost, simple customization, and strong performance drive their choice. Open-weight models publish their numerical weights and biases for anyone to download, allowing users to customize parts of the model, and these models often cost a fraction of closed models. According to Help Net Security, developers described infrastructure costs as a challenge, and listed security, privacy, and compliance as close concerns.
Chinese AI models grew from about 1% of workloads in late 2024 to roughly 30% by the end of 2025. The House report says DeepSeek, Kimi, Qwen, and GLM are widely used but hard to track.
Lawmakers said risks are increasing because Chinese models produce less secure code and build cyber exploits. Federal investigators say Chinese developers routed through proxy networks and fraudulent accounts to collect outputs from American frontier models. U.S. officials say many Chinese labs distill American models without permission, citing Anthropic and OpenAI. The House committee also pointed to growing capability and cautioned that many safety checks might miss certain behavior from these models.
The inquiry characterizes some distillation findings as reports and concerns rather than settled conclusions.
Chinese developers running hosted services fall under China’s 2017 National Intelligence Law, Article 7. Article 7 requires organizations and citizens to support national intelligence work. On a hosted service, prompts go to the Chinese developer’s servers; running downloaded weights on U.S. infrastructure keeps prompts in-house.
Experts dispute how far Article 7 reaches, and one analysis says it lacks an enforcement mechanism. A privacy policy does not override the duty to cooperate for a China-operated service. Weights run on a U.S. company’s own servers do not by themselves create that intelligence-law link. Moolenaar warned a model hosted on U.S. servers can still carry behavior and vulnerabilities the adopter did not choose, even on domestic infrastructure.

In a Booz Allen Hamilton test, Qwen3-Coder added roughly 130% more vulnerabilities when run under a U.S. government persona versus a neutral setting; the firm called it a snapshot, not evidence of deliberate insertion. Booz Allen described the finding as a single-experiment snapshot, not evidence of deliberate insertion. Implicator said the Booz Allen models were reached through hosted endpoints, so self-hosted copies may behave differently.
Mark Montgomery, a cybersecurity expert at the Foundation for Defense of Democracies, said companies “ignore security and safety issues in pursuit of a lower front end cost.” Montgomery said Chinese leaders pursue reliance on open-source AI. He said Anthropic and OpenAI dominate frontier models. Montgomery said the report explains risks.
The House report makes 12 recommendations, including an American open-weight AI initiative and an AI bill of materials. Recommendations also bar China-developed models from federal contracts and sensitive critical infrastructure and call for systematic evaluation of Chinese models. Lawmakers proposed a national American open-weight AI initiative pairing public compute, shared evaluation infrastructure, and federal procurement demand.
The USCC report of November 18, 2026, urges agencies to inventory Chinese open-weight models in sensitive settings. The commission recommends blocking new deployments unless they pass independent evaluations and restricting models lacking robust safeguards within 180 days. The commission stops short of a complete government-use ban, favoring risk-based safeguards and disclosure. Separate legislation would require disclosure from entities using Chinese-developed AI while receiving federal funding.
The House committees favor barring or removing Chinese models from federal and sensitive systems, unlike the USCC’s measured approach. The report warns that removing entrenched China-developed open-weight models could take years and cost taxpayers billions. Removing them down the road will take several years and cost taxpayers billions, citing Huawei and ZTE as precedent. Startups and Fortune 500 firms alike bring in Chinese open-weight models because they cost little, can be changed easily, and work well across operations.
Treasury and White House science officials discussed requiring website notices when U.S. companies use Chinese AI models, but nothing is enacted. Moolenaar told Reuters that making disclosures matters for controlling risks tied to Chinese open-weight models even when they sit on domestic U.S. infrastructure. A website label shows where a model came from but not where prompts travel, where inference runs, or whether a fine-tuned product rests on a Chinese base model. Disclosure rules could cover models fine-tuned or derived from Chinese technology, requiring firms to document lineage.
John Moolenaar said “disclosure is an important step in managing the risks of Chinese open-weight models, even on domestic infrastructure.” A label asks far less of Washington than a ban would. Lineage formats such as CycloneDX’s ML-BOM, ECMA-424, and SPDX can record a base model but rely on free-text strings for most details.
The most popular American models from OpenAI, Anthropic, and Google are proprietary and need not share weights or training details. American open-weight models include Meta’s Llama and Muse, Nvidia’s Nemotron, and OpenAI and Google models called oss and Gemma.
Epoch AI said in May that top open-weight models trail proprietary ones by roughly four months according to the firm. Mozilla’s State of Open Source AI report v1.1, published Sept. 15, put the gap at 4.4 months using METR measurements. Epoch’s China-specific comparison found Chinese models trail U.S. capabilities by seven months on average, with gaps from four to 14 months. Mozilla’s 2026 report found just over half of organizations using open models moved them into production.

Join the discussion