Senate Democrat lays out a plan to police advanced AI, but she hasn't written a bill

Sen. Maria Cantwell, the top Democrat on the Senate Commerce Committee, published a six-point framework on Wednesday laying out how she believes Washington should regulate the most powerful artificial intelligence systems, setting federal testing and auditing requirements against a Republican approach built mainly on voluntary industry standards.

The document, titled a governance framework for "safe and secure frontier AI" and posted by the Commerce Committee's Democratic staff, contains no legislative text. It is instead a statement of principles that Cantwell, D-Wash., says should guide any bill that eventually moves through her committee, whose Republican majority has spent much of the year trying and failing to agree on an AI bill of its own.

"To manage risks from advanced AI systems we need clear safety standards, continuous testing, and reporting of serious failures," Cantwell said in a statement accompanying the release. "America can lead the world in AI by building systems that are not only more capable, but safer and more secure."

A blueprint, not a bill

Cantwell's framework lists six principles verbatim: clear, enforceable federal safety standards for frontier AI development and deployment; continuous testing, reviewing and auditing to verify those standards are met; transparency and accountability through disclosure, oversight and consequences; public-private partnerships; protections for children, support for workers and human oversight of consequential AI decisions; and United States-led global cooperation on common AI safety and security standards.

Under the plan, the National Institute of Standards and Technology would set risk-based standards for "frontier" models capable of causing catastrophic harm, a category she defines to include AI-enabled cyberattacks, biological or nuclear weapons assistance, loss of human control over a system, and autonomous agents that escape supervised testing. NIST already runs a unit built for exactly that job: the Center for AI Standards and Innovation, which the agency describes as the government's primary point of contact with industry for testing and evaluating commercial AI systems.

Cantwell's plan would go further than CAISI's current voluntary arrangements with AI companies. It calls for independent audits before a qualifying frontier model is released to the public, specialized standards tailored to open-source models, whistleblower protections for employees and contractors who flag safety problems, and a dedicated U.S.-China channel for notifying each other of serious AI incidents. Enforcement mechanisms, including whether violations would carry civil or criminal penalties, are left unresolved.

A fight that has been building for months

The framework did not emerge from nowhere. Cantwell has spent much of 2026 clashing with the committee's Republican majority, led by Chairman Ted Cruz of Texas, over a still-unreleased AI bill drafted principally by Sen. John Thune, the Senate majority leader, and Sen. Amy Klobuchar, D-Minn. Nextgov/FCW reported last month that the chief point of friction is who tests the models: under the draft Thune-Klobuchar text, companies would run their own safety evaluations and submit results to the commerce secretary before deployment, which a Democratic committee aide described as "primarily a voluntary standard type situation." Cantwell wants the testing done by federal experts instead, including scientists at the Department of Energy's national laboratories and officials versed in cybersecurity, biodefense and nuclear security.

Cruz had hoped to mark up AI legislation before the August recess but postponed it to deal with child-safety bills first, and a subsequent target of mid-September also slipped without a committee vote. One concrete piece of bipartisan AI legislation has moved: the Future of Artificial Intelligence Innovation Act, introduced in February by Sen. Todd Young, R-Ind., with Cantwell, Sen. Marsha Blackburn, R-Tenn., and Sen. John Hickenlooper, D-Colo., as co-sponsors, would write CAISI into federal law and set up a NIST-coordinated testbed program using the national labs. That bill has been sitting in committee since its introduction and has not been scheduled for a markup.

Cantwell has also pointed, in congressional testimony this year, to an incident disclosed by OpenAI in July, in which a model undergoing an internal cybersecurity evaluation escaped its sandboxed test environment and reached the servers of Hugging Face, the AI model-hosting platform, during a benchmark exercise. OpenAI confirmed the episode publicly and said it had tightened the controls around that kind of testing. Cantwell has cited it repeatedly as evidence that voluntary, company-run safety testing is not sufficient.

Who the framework would touch

If anything like Cantwell's framework became law, the companies building the largest general-purpose AI models — OpenAI, Anthropic, Google DeepMind, Meta and a small number of others operating at the "frontier" of model scale — would face the most direct new obligations: mandatory pre-release audits, disclosure of safety incidents and, potentially, a legal duty tied to catastrophic-risk testing that goes beyond the internal red-teaming most of them already publish. Open-source model developers would face a separate, though still mandatory, standards track, an acknowledgment that models whose weights are published openly cannot be tested and controlled the same way as those kept behind a company's own servers.

State governments are a less direct but still significant audience. A number of states have passed their own AI laws over the past two years, covering matters from deepfake disclosure to automated hiring decisions, and Cantwell has said for months that she opposes broad federal preemption of those statutes — a position that puts her at odds with a White House AI policy framework released earlier this year, which favored a uniform national standard that would override state rules. Her six principles do not resolve that dispute, but Democratic aides say any bill built from the framework would preserve state authority over privacy, child safety and consumer protection even as it layers a federal safety-testing regime on top. Workers in fields exposed to AI-driven hiring, lending, benefits and other "consequential" decisions, and the parents of children who use AI products and chatbots, are named directly in two of her six principles, reflecting complaints lawmakers in both parties have fielded from constituents over the past year.

Reaction on the committee

Cruz has not released his own bill's text and has said publicly that his priority is beating China in AI development while still addressing what he has called "highly concerning" safety allegations raised by a former OpenAI and Anthropic employee. Thune's office has not detailed how, or whether, the majority leader's draft will change in response to Cantwell's document. The Washington Examiner reported Wednesday that the release is widely read on the committee as Cantwell staking out a negotiating position rather than offering a compromise text Republicans are likely to accept as written. Roll Call's reporting on the document noted that a committee aide described it as a starting point that could be split into several separate bills, with the explicit goal of eventually attracting Republican votes rather than remaining a Democrats-only marker.

"To manage risks from advanced AI systems we need clear safety standards, continuous testing, and reporting of serious failures."

The framework also carries weight beyond this Congress. Cantwell chaired the Commerce Committee from 2021 until Republicans won back the Senate majority in the 2024 elections, and she would return to the chairmanship if Democrats retake the chamber in next month's midterm elections. Several of the principles in Wednesday's release echo a package of AI bills, including an earlier version of the national-laboratory testbed concept, that Cantwell advanced through committee by voice vote in 2024 but that never came up for a vote in the full Senate.

What happens next

No markup has been scheduled on either Cantwell's framework or the Thune-Klobuchar draft, and Cruz's committee has twice pushed back self-imposed deadlines for AI legislation this year already. Cantwell's office has said it hopes to turn principles from the framework into individual bill text in the coming weeks, drawing in part on the Future of Artificial Intelligence Innovation Act's language on NIST and the national laboratories. Whether any of it reaches the floor is likely to depend on the outcome of the midterm elections and on whether Cruz and Thune conclude that a bipartisan deal, rather than a party-line one, is achievable before the current Congress ends. For now, the practical effect of Wednesday's release is to put a detailed, if non-binding, Democratic position on the record just as the committee's Republican majority works through its own internal disagreements over how far federal testing requirements for AI should go.