Skip to main content
v2026.11,986 entries · CC-BY 4.0

Editorial · CASRAI · AI and ML research outputs

House Trio Releases SAFE AI Act Outline for Frontier AI Safety Rules

Reps. Beyer, Jacobs and Maloy released a bipartisan outline requiring frontier AI developers to publish safety frameworks and pass evaluations.

Published 19 Nov 2026· 7 minute read
The U.S. Capitol building in Washington, D.C., seen in the evening from Pennsylvania Avenue
Photo: NASA/Bill Ingalls / NASA Image and Video Library, Public domain (cropped). Source · Licence

On 19 November 2026, US Representatives Don Beyer (D-VA), Sara Jacobs (D-CA) and Celeste Maloy (R-UT) released a two-page outline of the Safe and Accountable Federal Evaluations of Artificial Intelligence (SAFE AI) Act. It would require frontier AI developers to publish safety frameworks, report incidents and pass pre-deployment evaluations, and it says startups and academics are not generally covered. It is an outline, not bill text.

What happened

The three House members published the outline on 19 November 2026 and described the SAFE AI Act as “forthcoming bipartisan legislation that would create federal safety requirements for frontier artificial intelligence.” Their joint press release says the bill “would mandate safety frameworks, incident reporting, and increased transparency from frontier labs with strong government oversight and due process.”

Bloomberg Government reported the same day that the trio “outlined a sweeping new artificial intelligence proposal” that would require companies to meet minimum federal safety standards, disclose lapses and clarify where liability lies for “foreseeable injuries” tied to the technology. It described the announcement as previewing a bill that has not yet been introduced.

Rep. Beyer said the sponsors are “continuing to seek input on our forthcoming legislative text from experts and safety leaders, in coordination with committees of jurisdiction,” and called the outline a blueprint meant to encourage “further constructive feedback and discussion.”

Key details of the outline

The outline groups the proposal into three parts: safety frameworks and disclosures, independent evaluations with government oversight, and enforcement.

Area What the outline says
Safety frameworks Frontier developers must publish and adhere to enforceable frameworks stating what they test for, what thresholds trigger mitigations and what those mitigations are. Frameworks must be filed and published and “cannot be unilaterally loosened.”
Federal floor The Center for AI Standards and Innovation (CAISI), working with interagency partners including the intelligence community and the Department of Energy, writes minimum requirements for the frameworks. States can shape regulation above the federal floor.
Automated AI research and development Developers must disclose substantially autonomous AI research and development under prescribed measures.
Coverage Scope, risk domains and requirements can evolve over time. “To avoid stifling innovation, startups and academics are not generally covered.”
Incidents Any person may confidentially report a safety incident to CAISI, and developers must report safety incidents. Evaluators and developers must also report imminent risk of severe harm.
Evaluation Government teams evaluate each developer’s framework. Third-party verifiers, accredited by the government for specific domains, share assurance work while government teams keep ultimate authority.
Evaluator independence Developers do not pay evaluators directly, and there are limits on funding from one covered developer and its affiliates. CAISI assigns evaluators by qualification with mandatory rotation, and conflict-of-interest rules restrict movement of staff between developers and accredited evaluators.
Pre-deployment Evaluations assess framework compliance before commercial release or broad internal use, including automated AI research and development.
Enforcement Narrow emergency authority for the government to order mitigation where safeguards are inadequate and an AI system poses imminent risk of severe harm, subject to prompt judicial review. Civil penalties for framework violations, with remediation and judicial review, enhanced penalties for knowing or willful violations and procedures for criminal referral.
Liability and whistleblowers Clarifies developer liability for reasonably foreseeable injuries to third parties. People inside covered entities who raise safety concerns are protected.

The outline’s background section argues that frontier developers write their own safety rules, that no one outside the company can reliably verify compliance, and that AI systems deployed inside the companies, including to build the next generation of AI, receive little outside scrutiny.

Background and context

The outline arrives after several earlier federal and state moves. In June 2026, GovInfoSecurity reported on the Great American Artificial Intelligence Act, a 269-page discussion draft from Reps. Jay Obernolte and Lori Trahan. That draft would bar states from enforcing laws specifically regulating the development of AI models for three years, codify CAISI in statute with $100 million a year for fiscal 2027 to 2029, and apply to large developers with more than $500 million in revenue. Bloomberg Government reported on 23 July 2026 that the two members unveiled the FRONTIER Act to increase federal oversight of advanced models. The SAFE AI outline takes the opposite position on states: its authors call it “a floor – not a ceiling.”

At state level, California’s SB 53, the Transparency in Frontier Artificial Intelligence Act signed in 2025, requires frontier developers to publish safety frameworks, report specified critical incidents and protect whistleblowers, according to the governor’s office. On 18 November 2026, one day before the SAFE AI outline, California released recommendations for state-set baseline standards, ongoing third-party oversight and emergency shutdown authority; CASRAI covered them in its report on the California safety recommendations. New York’s RAISE Act has a registration deadline of 1 January 2027 for large frontier developers, as described in CASRAI’s RAISE Portal report.

CAISI is the body the outline relies on most. For what the centre is and how its testing works, see what CAISI is and how its pre-deployment testing compares with the UK’s. CASRAI also reported on Associated Press reporting about CAISI staff and the White House, relevant because the outline would give the centre new duties.

Reaction

Rep. Jacobs said: “No industry gets to police itself, and AI shouldn’t be the exception,” adding that the proposal “is built to adapt as the technology changes, so Americans’ safety never depends on a company’s promise.” Rep. Maloy said the bill “gives them clear rules, brings in an outside set of eyes to check their work, and leaves room for Utah and other states to set safeguards of their own.” Rep. Beyer described the outline as “a roadmap towards a safer AI future for the country and the world.”

California’s 18 November release, issued the day before, also called on Congress to adopt California’s framework or use it as a floor.

What it means for researchers, universities and research offices

Nothing in the outline is law, and no university has a compliance step today. The points below are editorial reading of the outline, not statements from the sponsors.

  • Academic AI labs are largely outside the proposed duties. The outline says academics are “not generally covered.” Obligations would fall on frontier developers, so a university lab training its own models would not be the target. How “frontier developer” would be defined is for the bill text.
  • Collaboration with covered developers is where exposure may arise. Sponsored research, model-access agreements and joint projects with a covered developer could pick up that developer’s framework, incident-reporting and evaluation obligations. Research offices may want to know which of their agreements involve frontier developers.
  • An accredited-evaluator market could create work for academic groups. The outline envisions government-accredited third-party verifiers accredited “in their accredited domains.” Universities with evaluation expertise could be candidates, but the outline sets no accreditation criteria.
  • Funding and independence rules would bind evaluators. Developers would not pay evaluators directly, with limits on money from any one developer and affiliates, and conflict-of-interest rules would restrict movement between developers and evaluators. Groups that both receive industry funding and evaluate models should watch how these are drafted.
  • Researchers could report incidents. Any person may confidentially report a safety incident to CAISI, and whistleblowers inside covered entities are protected. A researcher who finds a dangerous model behaviour would have a federal channel.
  • Federal floor, state ceilings. Because states could go further, institutions in California and New York would still track those laws alongside any federal bill. See the frontier AI safety hub for the wider landscape.

What happens next

The sponsors say legislative text is forthcoming and that they are seeking input from experts and safety leaders in coordination with committees of jurisdiction. No introduction date, bill number or hearing has been announced. California’s recommendations are expected to inform its legislature when it returns in December 2026, per CASRAI’s earlier report. Watch for bill text defining who is a frontier developer, how third-party verifiers are accredited and how the outline’s federal floor interacts with the Obernolte-Trahan proposals.

Sources

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →