A Senate Draft Would Make Frontier-AI Safety a Legal Duty — and Let Washington Block a Release

On 11 September 2026 Reuters reported that Senate negotiators were drafting a bill that would impose a legal duty of care on the developers of the most capable AI models, give the federal government power to block the release of a model it judges unsafe, and stop states enforcing their own laws on the same catastrophic risks. Courtney Rozen's report rested on two Senate aides, a person familiar with the deliberations and a lobbyist; syndicated copies carried it on 12 September, the date the archive uses. No bill text had been published, nobody outside the talks had seen one, no bill number existed, and no vote was scheduled. No model was withdrawn and no company was penalised — what the draft would change is who decides whether a frontier system ships. The archive grades this record B — documented, but from a single or secondary source — because every account of the proposal traces back to that same body of aide-sourced reporting. The full record, with its severity and confidence grades, is held in the Orca AI Incident Archive, alongside the neighbouring US governance entries from those weeks.

What the draft would do, on the aides' account

Four provisions recur across the reporting. Each was described to Reuters by people in the talks, not quoted from a text.

• A duty of care. Developers would have to design their products "with the goal of preventing catastrophic risks," as Reuters put it. The concept is borrowed from tort law, where an actor owes a duty to take reasonable precautions against foreseeable harm. Tech Times, in its 12 September write-up, described the practical consequence as an obligation baked into the design phase rather than one enforced only through audits and incident reports, so that liability could attach to the design decision itself and not merely to a failure to file a report. Courts, not only regulators, would end up deciding what counts as reasonable.

• A federal power to block a release. The government could block the release of "certain models judged unsafe," and a company could challenge that decision in federal court. How much authority the executive would hold was, and remains, under discussion.

• Independent testing. The proposal would involve government experts, including the national laboratories, in assessing whether frontier models could enable sophisticated cyberattacks or assist in developing biological or nuclear weapons. Klobuchar said on X that any agreement has to make developers "work with government experts to verify and test models to make sure AI is safe."

• Preemption of state law. Part of the bill would stop states enforcing their own laws on certain model risks. One Senate aide cited the scenario of a bad actor using an AI system to help design nuclear or biological weapons.

The three risk areas named across the reporting are assistance with biological or nuclear weapons, the ability to run sophisticated cyberattacks against critical infrastructure, and misuse by determined bad actors. The bill would apply only to "models with the most advanced capabilities." Reuters identified Alphabet's Google, Anthropic and OpenAI as among the US companies holding models in that class. That is the wire service's identification of who would be covered, not a list of covered developers written into any text — because there is no public text.

Who is in the room — and who is at the door

Reuters named three negotiators: Majority Leader John Thune of South Dakota, Commerce Committee chairman Ted Cruz of Texas, and Amy Klobuchar of Minnesota, the senior Democrat involved. A Thune spokesperson declined to comment. Cruz wrote on X that he was working with Klobuchar and Thune "to address catastrophic risks involving biological or nuclear threats." Klobuchar framed the effort around government verification and testing.

The fourth name needs care. The archive record lists Maria Cantwell, the ranking Democrat on Commerce, among the bipartisan group drafting the bill. The reporting places her differently: outside the negotiating trio, and the central obstacle inside it. The Next Web's reading of the Reuters and Nextgov reporting is that Cantwell "is blocking over the testing framework," because she wants models sent to federal agencies and the national laboratories rather than tested by the companies that built them. Her office offered alternative language to Klobuchar's, which an aide described as receptive, and by that account Cruz's approval is the deciding factor. Reuters' follow-up on 14 September says only that Cantwell "has also weighed in on the legislation, sources say." Her participation is established; whether she is a drafter is not.

The same reporting names the outside parties. A second source told Nextgov that Cantwell, Anthropic and unnamed AI-safety groups were "refusing to play ball." Anthropic did not answer Nextgov's request for comment.

The machinery: who actually runs the tests

This is where the draft's headline and its plumbing diverge, and it is the part most summaries drop. According to the Reuters reporting as summarised by The Next Web, "Companies would run the safety tests themselves, then present the results to the Commerce secretary for deployment approval." A Democratic committee aide characterised the resulting arrangement as a "primarily a voluntary standard type situation."

Two days after the record date, Reuters reported that negotiators were also weighing whether the commerce secretary could "request proof that AI developers are taking reasonable steps to prevent harm," and whether the secretary could "deploy government auditors to test companies' products themselves." The same report conceded that Reuters "could not immediately determine what would qualify as 'reasonable.'" That gap is the whole argument. A duty of care is only as strong as the standard a court or a regulator applies to it, and the standard was still being written.

Accounts of the gate itself have varied since. A 14 September round-up on eworld.live, drawing on Reuters and Nextgov, described the open question as whether Commerce gets "a true pre-deployment block vs softer attestation," and called that contested. Both descriptions can be true of the same draft at different stages of negotiation; neither can be verified against a text that does not exist.

Cantwell's red line, and Cruz's answer

Cantwell's public position, as Reuters reported it, is that meaningful legislation would send the most powerful models to scientists at the national laboratories, who would "assess whether they could enable sophisticated cyberattacks or aid the development of biological or nuclear weapons." In her account the testers are federal, not corporate — the point of the exercise being that a company's own safety case is not evidence.

Cruz's answer came five days after the record date, and it sharpens the fault line. The Houston Chronicle reported on 17 September, in a piece by James Osborne, that Cruz was working to block a Democratic proposal for mandatory federal pre-deployment testing, warning it "would hurt U.S. companies and give a leg up to Chinese competitors." His alternative was a requirement that companies "take steps to prevent their models from developing 'advanced threat capabilities'" and "disclose to the federal government any active or imminent security breaches." He also argued that market liability already does the work: "If any company sells a product that causes massive harm, they'll get sued into oblivion."

That reporting is later than the record and is presented here as what it is — a later development that locates the September draft rather than revising it. The proposal reported on 11 September gave the government a blocking power and spoke of testing by government experts. The chairman of the committee that would have to mark it up was, within a week, arguing against compulsory pre-deployment federal testing.

Preemption, and the impossibility of naming the statutes

The preemption clause draws the clearest opposition and the least clarity. Reuters reported that part of the bill would stop states enforcing their own laws on certain model risks. Which laws those are is where the coverage comes apart: The Next Web names California's SB 53 and New York's RAISE Act, while Tech Times reports that California signed SB 813 and AB 1405 on 9 September 2026 and describes California's mandatory-audit law as covering AI used in hiring, insurance and other consequential decisions. This article did not review the chaptered texts of either state's statutes, so which laws the clause would displace is unresolved in the sources rather than settled by them.

What is clear is the history behind the fight. Tech Times dates an earlier blanket moratorium on state AI regulation to 2025 and records its defeat in the Senate by 99 votes to 1 after Cruz pushed for it; AIBARS describes a similar effort in 2026 being stripped out by an overwhelming vote. The two accounts do not reconcile on the year, and no roll-call record was reviewed for this piece. Tech Times adds that the open question is whether a narrower preemption confined to catastrophic-risk obligations can survive objections from lawmakers in both parties who treat state rules as a consumer-protection backstop. AIBARS, for its part, reports that Republican governors in several states have opposed broad preemption, and that civil-liberties and consumer groups warn a federal standard could become a ceiling rather than a floor.

Established, attributed, and not established

• Established across independent outlets. That senators are negotiating a duty-of-care bill with a federal blocking power and a preemption clause is reported by Reuters on 11 and 14 September, by Nextgov on the testing dispute and by Semafor; the Houston Chronicle reports the Cruz side of the fight rather than the bill's contents. No outlet has published the text, because no text is public.

• Established: the split inside the negotiation is real and reported twice from different directions. Nextgov reported Cantwell's alternative language and her preference for federal testing; the Houston Chronicle reported Cruz working against mandatory federal pre-deployment testing. Those are two named offices, two reporters, and the same fault line.

• Attributed only. Every operational detail — the trigger for coverage, what makes a model "unsafe," whether Commerce can block a release or only demand an attestation, which developers are in scope — comes from unnamed Senate aides, a person familiar with the talks and a lobbyist. Thune's office declined to comment. Cruz and Klobuchar addressed the effort only in general terms on X. Tech Times also quotes a Democratic committee aide calling the draft "primarily a voluntary standard type situation," which sits awkwardly beside the "binding duty of care" framing and is not a contradiction so much as a description of the same document from a different seat.

• Not established: that the bill will be introduced at all. No bill number, no committee print, no scheduled markup and no vote as of the reporting reviewed. Reuters' own assessment was that the measure "faces long odds of becoming law."

• Not established: whether Cantwell is a drafter. The archive record lists her in the group; the reporting places her outside the core three and in conflict with them over testing. The discrepancy is small and worth flagging, because her name changes what the bill's coalition looks like.

The calendar, the courtroom, and what would happen next

The schedule is the draft's most reliable opponent. Reuters reported that the House was due to sit for one week and the Senate for about three before the 3 November midterms, after which a new Congress arrives in January 2027. Cruz had said the committee might hold a markup; an earlier markup slipped before the August recess when he prioritised child-safety legislation, Nextgov reported. Tech Times quoted Semafor's sources calling the measure "the only viable option to stand a chance before 2027," a line that reads as much like a deadline warning as an endorsement.

The White House is not a reliable ally either. On 14 September Reuters reported that President Donald Trump argued existing authorities already provide adequate tools to regulate and prosecute technology companies — a position that signals reluctance to sign a bill creating new AI rules. Thune, for his part, framed the case in competitive terms, saying there is "a way in which Congress can put guardrails around those types of more consequential threats" while insisting that doing so would not damage the US position in the AI race.

The legal risk is the other constraint, and it is structural rather than political. Tech Times notes that a pre-release government block is a form of prior restraint, a category courts scrutinise closely, and that whether an AI model release counts as protected speech is unresolved; legal scholars told the outlet that a First Amendment challenge would be the primary challenge to any such statute. The same piece characterises the bill as the first purpose-built AI blocking statute with a court-appeals route, and contrasts it with earlier uses of existing authority — the Export Control Reform Act invoked in June 2026 to suspend global access to Anthropic's Mythos 5 and Fable 5 for about three weeks, and a 2 June 2026 executive order creating voluntary pre-release government access without any preclearance requirement. Those characterisations are Tech Times', and the underlying national-security actions were not independently reviewed here.

What the record is worth

The value of this entry is not that a bill exists. It is the distance between the sentence everyone repeated — a binding duty of care, a government that can block a release — and the machinery underneath it, in which the companies run their own safety tests, the results go to a cabinet secretary, and the aide closest to the drafting calls the result "primarily a voluntary standard type situation." Both descriptions are of the same document. One is the version that survives a headline; the other is the version that would survive a markup.

Three things would settle it, and none had happened in the reporting reviewed for this record: a bill number and a published text, a Commerce Committee markup, and a decision on whether the federal role is a veto or an attestation. Until then the four provisions should be read as a negotiating position described by people in the room, not as law in waiting — and the archive's B grade is the right label for it, because the only witnesses are anonymous and the document itself has not been seen.