Three members of the US House released an outline on November 19, 2026 for a bill that would create the first federal safety requirements for frontier AI developers. Reps. Sara Jacobs (D-CA), Celeste Maloy (R-UT) and Don Beyer (D-VA) call it the Safe and Accountable Federal Evaluations of Artificial Intelligence (SAFE AI) Act. Their press release says it "would mandate safety frameworks, incident reporting, and increased transparency from frontier labs with strong government oversight and due process."
The key fact for readers: this is an outline, not a bill. There is no legislative text, no bill number and no committee schedule yet. Here is what the outline says, how it fits the other proposals in Congress and what it means if it advances.
TL;DR: the SAFE AI Act outline
| Question | Answer |
|---|---|
| Who released it? | Reps. Sara Jacobs (D-CA), Celeste Maloy (R-UT) and Don Beyer (D-VA) |
| When? | November 19, 2026 |
| Is there bill text? | No. Beyer says they are still seeking expert input |
| Who is covered? | Frontier developers; startups and academics generally excluded (MLQ summary) |
| Core duties | Published safety frameworks, incident reporting, pre-release evaluations |
| Who evaluates? | Government teams and CAISI-assigned, government-accredited third parties |
| Enforcement | Civil penalties, higher for willful violations, criminal referrals, emergency orders |
| States? | A federal floor that states can build on, per the sponsors |
What does the outline say?
The sponsors' press release is short on mechanics, so the details below come from MLQ's summary of the two-page outline, which cites the PDF. Check the outline itself before relying on any single line.
Safety frameworks. Each frontier developer would file and publish a framework stating what it tests for, which thresholds trigger mitigations and what those mitigations are. The Center for AI Standards and Innovation (CAISI) at the Commerce Department would write minimum requirements with partners including the intelligence community and the Energy Department. Developers could not loosen a framework unilaterally.
Autonomous AI research. Developers would also have to disclose "substantially autonomous" AI research and development. That reaches the question of models doing AI research with little human input, which we covered in Hickenlooper on recursive self-improvement.
Incident reporting. Anyone could confidentially report a safety incident to CAISI. Developers and evaluators would have to report incidents and imminent risks of severe harm.
Evaluations. Government teams would review each framework. Government-accredited third parties would run compliance evaluations before commercial release or broad internal use. CAISI would assign and rotate evaluators, paid from fee-based funds and not by developers directly, with limits on how much funding can come from one company.
Enforcement. Civil penalties, higher penalties for knowing or willful violations and criminal referrals. Where safeguards are inadequate and a system poses imminent risk of severe harm, the government could order mitigations, subject to prompt judicial review. The outline would also clarify developers' liability for reasonably foreseeable injuries to third parties.
How would the process work for a lab?
Columns forming a small government building with a green flag, standing for federal oversight of frontier AI labs under the SAFE AI Act outline
Reading the outline as a sequence helps. A frontier lab would first publish a framework that says what it tests for and what thresholds trigger mitigations. Government teams, working with CAISI, would review that framework against CAISI's minimum requirements. Before a model goes into commercial release, or into broad internal use, a government-accredited third party would run a compliance evaluation. CAISI would pick and rotate the evaluator, and the lab would pay into a fee-based fund instead of hiring the evaluator itself.
That last design choice answers a common criticism of today's voluntary testing, where a lab chooses and pays its own auditor. Rotation and funding caps are meant to reduce that pressure. They also create work: accrediting evaluators, setting fees and avoiding delays for releases.
If something goes wrong, two routes open. Anyone can report an incident confidentially to CAISI, and labs and evaluators must report incidents and imminent risks. If a system poses imminent risk of severe harm and safeguards fall short, the government could order mitigations, and the lab could seek prompt judicial review. The sponsors describe this as "targeted emergency authority" with "due process."
Questions the outline does not yet answer
| Question | Why it matters |
|---|---|
| What is a frontier model? | The threshold decides who is covered; no number appears in the summaries we read |
| What is "broad internal use"? | Labs often run models internally long before release |
| How long would evaluations take? | Delays affect product schedules and competition with foreign labs |
| Who pays if fee funds fall short? | CAISI capacity is a practical limit |
| How do state laws such as New York's RAISE Act interact? | A floor leaves state rules in place |
What did the sponsors say?
Jacobs: "No industry gets to police itself, and AI shouldn't be the exception." She added that "companies write their own rules while their own models break out of test environments and hack other companies' systems," and called the plan "a floor, not a ceiling, empowering states to adopt stronger protections."
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."
Beyer called it "a roadmap," 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 said Congress should "tackle AI safety swiftly, and that we get it right." The incident Jacobs alludes to is the kind of agent-escape case we covered in the Arizona criminal probe into OpenAI over rogue bank agents.
How does it fit with other federal and state efforts?
The House already has a growing set of AI safety efforts, which is why a three-member outline matters mainly as a signal of where bipartisan support may gather.
- The House AI caucus. We covered the AI Safety and Innovation Caucus, where Beyer is a member, as the group building cross-party consensus.
- State rules. MLQ notes that New York has opened RAISE Act registration for frontier labs, and a California report this week proposed emergency shutdown powers. See our post on California's AI safety framework and on Scott Wiener's move to Congress.
- Shutdown and fail-safe ideas. The Paris call for fail-safe shutdowns points the same direction as the outline's emergency orders.
- The White House. MLQ points out the outline lands the same day Trump told APEC leaders the US opposes "globalist" regulation of super intelligence, which we cover in Trump and Xi on AI at APEC. Opposition to global rules is not the same as opposition to a domestic bill, but the administration has not said it supports this one.
What is still unknown
- Bill text and number. Until text exists, details such as what counts as a frontier model are unknown. The summaries we read give no compute or capability threshold.
- Preemption. The sponsors call the proposal a floor, so state laws would stay. The outline's exact language was not available to us, so check it for how federal and state duties would interact.
- Committee path. Beyer says the sponsors are coordinating with committees of jurisdiction. No markup or hearing has been announced.
- Industry and agency reaction. The coverage we reviewed does not include responses from labs or from Commerce.
- Funding and capacity. Government-run evaluations need staff and fee-based funds. Whether CAISI can do this at scale is a practical question the outline does not answer.
What this means for what you build or pay
For builders at startups and in academia, the reported carve-out means most small projects would not be covered. For anyone who builds on a frontier lab's API, the effect would be indirect: slower releases if pre-release evaluations take time, and more published documentation about model risks.
For founders, a federal floor that states can raise means a lab could face federal and state duties together, so build compliance around the strictest rule that applies to you. For marketers and policy teams, describe the SAFE AI Act as a proposed outline. Writing that it "requires" anything would be wrong today.
Is this likely to pass?
Nobody can say yet, and an honest answer is that the outline is an early step. It has bipartisan sponsors, which helps. It also has no text, no committee path and no stated support from the administration or from the leaders who control the House floor. Many AI bills have been drafted and few have moved, so treat the outline as a statement of where a group of members want the debate to go. The useful signal to watch is whether more members of both parties sign on once text appears, and whether the labs respond with support, objections or their own proposals.
How to follow it
- Read the sponsors' press release and the linked outline PDF.
- Watch for a bill number on congress.gov, which would mean text was introduced.
- Watch whether committees with jurisdiction schedule hearings.
- Compare with our look at an independent test where chatbots produced attack plans, which shows why lawmakers want evaluation duties.
Related reading
- House AI Safety and Innovation Caucus
- California AI safety framework
- Trump and Xi on AI at APEC
- Paris call for fail-safe shutdowns
- Hickenlooper on recursive self-improvement
- Arizona probe into OpenAI rogue bank agents
- Scott Wiener and federal AI regulation
Details are as stated by the sponsors and reported by MLQ on November 19, 2026. An outline can change before any bill is introduced.
