House Lawmakers Urge US Open-Weight AI Initiative, Citing Risks in Chinese Models

Summary
A House committee report says Chinese open-weight AI models are widely used and warns they may produce less secure code and autonomously build cyber exploits. Lawmakers recommend U.S. models, evaluations and limits on federal use.
Key points
- The House Select Committee on China said Chinese open-weight models are widely used, though much of that use is difficult to track.
- Organizations cited low cost, customization and performance as reasons for using models including DeepSeek, Kimi, Qwen and GLM.
- The committee warned that Chinese models may produce less secure code and autonomously build cyber exploits.
- Lawmakers proposed a national U.S. open-weight AI initiative with public computing resources, shared evaluation infrastructure and federal procurement demand.
- The committee also recommended prohibiting federal agencies from using Chinese models and sanctioning companies that illicitly distill U.S.-made models.
- The report cautioned that some safety evaluations may not detect certain behavior in these models.
Article Details
- Event Type
- House Select Committee on China report and policy recommendations concerning Chinese open-weight AI models
- Impact
- The committee reported that Chinese open-weight models are widely used and warned that they can produce less secure code and autonomously build cyber exploits. It recommended a U.S. open-weight AI initiative, restrictions on federal agency use of Chinese models, consistent model evaluations, and sanctions for illicit distillation of U.S.-made models.
Vendors
AnthropicThe most popular American-made models, those from OpenAI, Anthropic and Google, are all proprietary models, meaning these companies are not required to share the weights of their models or release any detailed trainingAnysphereare. The report is based on an investigation begun in April into why three American companies, Airbnb, Anysphere, Cursor and DoorDash, turned to these models.GoogleThe most popular American-made models, those from OpenAI, Anthropic and Google, are all proprietary models, meaning these companies are not required to share the weights of their models or release any detailed trainingMetaSome American open-weight models do exist. Meta's Llama and Muse models have always operated under an open-source license and are considered open-weight. Nvidia, long an advocate for open-source AI, distributes theMoonshot AIChinese models. The company built its coding model Composer 2 using Kimi K2.5 from Chinese AI lab Moonshot AI.NvidiaLlama and Muse models have always operated under an open-source license and are considered open-weight. Nvidia, long an advocate for open-source AI, distributes the open-weight Nemotron. OpenAI and Google also haveOpenAIThe most popular American-made models, those from OpenAI, Anthropic and Google, are all proprietary models, meaning these companies are not required to share the weights of their models or release any detailed training
Products
Claude Fablemodels are developing quickly may be due to distillation, where developers prompt a teacher model like Claude Fable and use its responses and reasoning chains as training data for their own models. Such practice isComposer 2which SpaceX recently bought, plans to stop using Chinese models. The company built its coding model Composer 2 using Kimi K2.5 from Chinese AI lab Moonshot AI.Cursorreport is based on an investigation begun in April into why three American companies, Airbnb, Anysphere, Cursor and DoorDash, turned to these models.DeepSeekCommittee on China published a report focusing on how widespread Chinese open-weight models such as DeepSeek, Kimi, Qwen, and GLM are. The report is based on an investigation begun in April into why three AmericanGemmaAI, distributes the open-weight Nemotron. OpenAI and Google also have open-weight models, oss and Gemma, respectively.GLMpublished a report focusing on how widespread Chinese open-weight models such as DeepSeek, Kimi, Qwen, and GLM are. The report is based on an investigation begun in April into why three American companies, Airbnb,Kimion China published a report focusing on how widespread Chinese open-weight models such as DeepSeek, Kimi, Qwen, and GLM are. The report is based on an investigation begun in April into why three AmericanKimi K2.5bought, plans to stop using Chinese models. The company built its coding model Composer 2 using Kimi K2.5 from Chinese AI lab Moonshot AI.LlamaSome American open-weight models do exist. Meta's Llama and Muse models have always operated under an open-source license and are considered open-weight. Nvidia, long an advocate for open-source AI, distributes theMuseSome American open-weight models do exist. Meta's Llama and Muse models have always operated under an open-source license and are considered open-weight. Nvidia, long an advocate for open-source AI, distributes theNVIDIA Nemotronand are considered open-weight. Nvidia, long an advocate for open-source AI, distributes the open-weight Nemotron. OpenAI and Google also have open-weight models, oss and Gemma, respectively.ossfor open-source AI, distributes the open-weight Nemotron. OpenAI and Google also have open-weight models, oss and Gemma, respectively.Qwenon China published a report focusing on how widespread Chinese open-weight models such as DeepSeek, Kimi, Qwen, and GLM are. The report is based on an investigation begun in April into why three American companies,
Countries
ChinaThe House Select Committee on China published a report focusing on how widespread Chinese open-weight models such as DeepSeek, Kimi, Qwen, and GLM are. The report is based on an investigation begun in April into whyUnited StatesHouse lawmakers recommended Friday that the Trump administration establish an initiative for American-made open-weight AI models that would entice organizations to turn away from Chinese open-weight models.
