House chairmen urge banning Chinese AI models from federal work, and Cursor's maker may drop them
The chairmen also warn that pulling the models before American ones exist would raise costs for the companies they examined.
On November 20, 2026, the chairmen of two House committees released a joint report asking Congress to keep AI models built in China out of federal systems, federal contract work and the riskiest jobs in critical infrastructure. Andrew Garbarino chairs the Committee on Homeland Security and John Moolenaar chairs the Select Committee on China.
The report rests on a seven-month investigation of three American companies: Airbnb, DoorDash and Anysphere, which makes the Cursor coding tool. It also carries one line that matters to a lot of programmers. Anysphere told the committees it plans to explore phasing out Chinese-developed open-weight models, citing geopolitical risk.
- Released
- November 20, 2026
- By
- Chairmen Garbarino and Moolenaar
- Companies examined
- Airbnb, Anysphere and DoorDash
- Recommendations
- 12 policy recommendations
- Investigation began
- April 2026
What the committees found inside three companies

The committees' release says Anysphere built the flagship proprietary model in Cursor on a Moonshot AI base model. The full report names it: Composer 2, built on Kimi K2.5, a fact the report says came out only after an outside developer spotted the underlying model. The phase-out line cites an Anysphere briefing and gives no timeline, and the report's wording is that Anysphere plans to explore it.
Airbnb runs customized versions of Alibaba's open-weight Qwen models in customer support, according to the report, including an assistant that handles about 80,000 guest and host conversations a day. DoorDash told the committees that 25 Chinese model families had been authorized and deployed there, and that swapping out one model built into a larger system could take up to six months.
The committees didn't stop at the companies. Their release says people tied to the People's Liberation Army's cyber unit visited the Beijing headquarters of Zhipu several times, most recently in August 2026, and that a July 2026 military procurement wanted DeepSeek built into a cyber training range.
What they want Congress to do
The headline recommendation extends the Homeland Security Department's existing ban on DeepSeek in federal civilian systems to every Chinese-developed model. The release lists it this way.
the Chairmen call for prohibiting PRC-developed models in federal systems and federal contract work, removing them from high-risk functions of critical infrastructure, and barring the use of federal cybersecurity grant funds to buy them.
The full report adds detail the release leaves out. The contractor ban would start with the largest contractors and reach the rest after a transition period, with extra time for small businesses. In critical infrastructure the ban would apply at once to new deployments, with removal schedules tied to when trusted alternatives exist.
The chairmen also want something built. Their first recommendation is a national American open-weight AI initiative, and the report's foreword spells out why. Forcing companies off Chinese models before American alternatives exist would, in the chairmen's words in the report, raise costs and push the market toward a few closed providers.
Who this reaches outside Washington
Federal contractors are the most direct group. If Congress adopts the contract rule, any company doing federal work would have to know which models sit inside its software, which is why the report also asks for an AI bill of materials, a list of the models a product is built from.
The bigger number is how far these models have already spread. The report cites OpenRouter data showing that the share of tokens from U.S. requests going to open-weight models rose from 26 percent in September 2025 to 60 percent in August 2026, with models from Chinese labs making up most of that volume. It also cites Hugging Face figures putting Chinese-developed models at 41 percent of downloads over the past year.
For developers the practical question is what happens to their tools. Cursor's users don't pick its base model, and the report doesn't say what Anysphere would move to or when. DoorDash's six-month estimate for replacing one embedded model gives a sense of the work involved. I'd read that as the real cost of the recommendations, and the chairmen's own foreword makes the same point.
The release says the committees will continue their oversight. Neither release names a bill to carry the recommendations.
More on AI policy
- Trump says American companies that build AI data centers abroad will face a 'tremendous tax'November 21, 2026
- Bessent says Trump and Xi agreed humans must stay in military AI decisionsNovember 20, 2026
- European Commission confirms a cyber campaign run with a public AI tool, detected October 22November 19, 2026
- Whitesides, Moran and Beyer launch what they call Congress's first AI safety caucusNovember 19, 2026


