US lawmakers propose security checks before powerful AI models reach the public

Two bipartisan House bills would require pre-deployment security testing for the most powerful AI models and ban federal agencies from using Chinese-developed open-weight systems. The proposals matter because they would move AI safety from voluntary promises towards a formal government checkpoint, with the usual small question of who gets to press the stop button.

Watch Desk analysis

What happened

Representatives Josh Gottheimer and Mike Lawler announced the American AI Security Act, which would require covered developers to give the National Security Agency access to assess whether a model could support serious cyberattacks or help create chemical, biological or radiological weapons. The proposed review would last 30 days, with one possible 30-day extension.

Gottheimer and Representative Nick LaLota also announced the China FIREWALL Act. The bill would bar federal agencies from using Chinese-developed open-weight AI models on government devices and from buying software that relies on them. It would extend an existing federal restriction on DeepSeek to models such as Moonshot’s Kimi K3 and Alibaba’s Qwen, according to Vital Law.

The article says the text of both bills was not immediately available. Gottheimer also urged House Speaker Mike Johnson to bring lawmakers back before the election to consider the measures. That is a political request, not evidence that either bill is on track to become law.

Why it matters

The proposed security review would test models before release rather than waiting for a harmful deployment or incident. Its scope is deliberately narrow, focused on national-security risks such as cyberattacks and weapons development, but it would still give a federal agency meaningful access to powerful systems.

The China FIREWALL proposal takes a different route: controlling which models and model-based software can enter government systems. That could reduce exposure to foreign-developed tools, while also making procurement more complicated and potentially excluding useful open technology. Open-weight models are software, not obedient filing cabinets, so defining what counts as “using” one may become an entertainingly thorny drafting exercise.

Our read

This is a credible and consequential policy development, but it is still a pair of announced bills rather than enacted law. The useful change for readers is the shape of the proposed choices: test high-risk models before deployment, restrict particular model origins in government, or continue relying mainly on existing rules and voluntary checks.

The strongest part is the attempt to attach a clear clock to security review. The weakest part, for now, is the missing bill text. Until lawmakers publish the language, questions about which models are covered, how the NSA would handle access and what happens after an adverse finding remain unanswered.

What to watch

• Whether the full text of both bills is published and how “most powerful” models are defined.
• Whether the proposed NSA review would be advisory, binding or capable of stopping a release.
• How the China FIREWALL Act would treat software that uses several models or is hosted outside China.
• Whether House leadership schedules hearings or a vote before the election. Sources and evidence: Vital Law, “AI Security Threats Prompt Bipartisan Legislative Push in House”, published 21 September 2026. The proposed bills, sponsors, review timetable, model examples and political response are attributed to that report and the lawmakers’ statements described in it. The bill text was not available in the supplied evidence.

Discussion spark: Should governments be able to delay the release of powerful AI models for national-security testing, or would that give security agencies too much control over a fast-moving technology?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.