A Senate draft would route frontier model launches through the Commerce secretary

The difference between the two versions on the table is who holds the model while it is being checked. In the draft a Democratic committee aide described to Nextgov, the developer runs its own safety tests and presents the results to the Commerce secretary, who approves deployment ^[11]. The company keeps the weights and adds a submission step whose duration nobody has published. Maria Cantwell wants the models sent to federal agencies instead, including the national laboratories and national security agencies ^[12]. In her version the model leaves the building. The aide called the draft "primarily a voluntary standard type situation" ^[13].

The blocking power would apply to models with the most advanced capabilities, and Reuters named Google, Anthropic and OpenAI as the US companies holding them ^[4]. The preemption clause reaches further. Two Senate aides told Reuters that part of the bill would stop states enforcing their own laws on certain model risks, one offering the example of a bad actor using an AI system to design nuclear or biological weapons ^[18]. Compliance work built against California's SB 53, the law OpenAI first opposed and then asked the state to toughen ^[19], and against New York's RAISE Act ^[20], sits in that path. Cantwell wrote on X that the clause is a sharp reversal from the attempt to impose a ten-year moratorium on state AI regulation ^[21]. The Senate stripped that moratorium from a bill by 99 votes to one in July 2025 ^[22].

Cantwell is the ranking Democrat on Commerce and is not in the negotiating trio, and Alexandra Kelley reported for Nextgov that she is blocking over the testing framework ^[10]. A second person familiar with the matter told Nextgov that Cantwell, Anthropic and AI safety groups are "refusing to play ball" ^[15]. Anthropic did not answer Nextgov's request for comment ^[16], and Thune's spokesperson declined to comment to Reuters ^[9]. Klobuchar said in a statement that any agreement has to make developers work with government experts to verify and test models, and that the US cannot allow the release of dangerous models, including those that evade the control of their developers ^[8].

Cruz said this month that the Commerce Committee may hold a markup, having cancelled one scheduled before the August recess because he wanted to focus on child safety legislation ^[7]. The midterms are on 3 November ^[24]. The Senate is expected in Washington for three more weeks, and the House is scheduled for one, returning on 9 November ^[25], six days after the vote ^[1]. Sam Liccardo drafted a letter asking the Speaker to bring the House back immediately and keep it in session until it advances bipartisan AI safeguards; George Whitesides, Lori Trahan and Ted Lieu co-signed it ^[26].

Federal review of the most capable models already happens under rules the public cannot read. TNW reported on 9 September that the White House has not released its framework, and that the companies who have read it are the largest ones ^[23]. A statute would at least produce a text a smaller developer could plan against.

The dependency sorts on two axes: who runs the test, and who decides the model ships. Today the company answers both. The negotiators' draft as described keeps the first answer inside the company and moves the second to the Commerce secretary ^[11]. Cantwell's alternative moves both ^[12]. The Democratic aide said Cantwell's office has offered language that Klobuchar's office is receptive to, and that Cruz's approval is the deciding factor ^[14].