Economy & Policy

House Trio Unveils SAFE AI Act Outline, First Federal Safety Rules for Frontier AI

Three House members propose the first federal safety rules for frontier AI: mandatory frameworks, third-party evaluations, incident reporting and civil penalties — with no bill text yet.

By Amara Osei

4 min read

Bipartisan House Trio Unveils SAFE AI Act Outline, Proposing First Federal Safety Rules for Frontier AI
Bipartisan House Trio Unveils SAFE AI Act Outline, Proposing First Federal Safety Rules for Frontier AIMohamed_hassan / pixabay

What's News

  • Reps. Jacobs, Maloy and Beyer released a two-page SAFE AI Act outline on Thursday, the first proposed federal safety requirements for frontier AI
  • The proposal mandates published safety frameworks, pre-deployment third-party evaluations, incident reporting and civil penalties
  • The bill sets a federal floor that explicitly allows states to impose stricter AI rules
  • The outline does not define 'frontier model' and has no bill text, number or markup schedule
  • It landed the same day Trump told APEC leaders the U.S. opposes 'globalist' AI regulation

Three House members released an outline Thursday for the first federal safety requirements governing companies that build the most advanced AI systems. Reps. Sara Jacobs (D-Calif.), Celeste Maloy (R-Utah) and Don Beyer (D-Va.) published a two-page framework for the Safe and Accountable Federal Evaluations of Artificial Intelligence (SAFE AI) Act — a bipartisan proposal, not yet legislative text, that would mandate safety frameworks, independent pre-deployment evaluations, incident reporting and civil penalties.

What does the outline actually require?

Under the proposal, "frontier developers" would have to publish safety frameworks detailing what they test for, what thresholds trigger mitigations and what those mitigations entail. The Commerce Department's Center for AI Standards and Innovation (CAISI), working with the intelligence community and the Energy Department, would write minimum requirements for those frameworks. Developers could not unilaterally weaken a framework once filed.

Before commercial release or "broad internal use" — including substantially autonomous AI research — a government-accredited third party would run a compliance evaluation. CAISI would assign and rotate evaluators, paying them from fee-based funds rather than by developers themselves, with caps on funding from any single company and conflict-of-interest restrictions on revolving-door hiring. Government teams would review each framework and retain ultimate authority.

The outline also mandates incident reporting:

  • Anyone could confidentially report a safety incident to CAISI
  • Developers and evaluators must report incidents and imminent risks of severe harm
  • Enforcement includes civil penalties, with higher penalties for knowing or willful violations, plus criminal referrals
  • A narrow emergency authority would let government order mitigations when safeguards are inadequate and a system poses imminent risk of severe harm — subject to prompt judicial review

The outline would also clarify developers' liability for reasonably foreseeable injuries to third parties. Startups and academic researchers would generally not be covered.

A federal floor, not a ceiling

The sponsors framed the bill as a baseline that preserves state action. "Our proposal is a floor — not a ceiling, empowering states to adopt stronger protections for their communities," Jacobs said. Maloy said the bill "leaves room for Utah and other states to set safeguards of their own."

The distinction matters. A GOP-led effort to preempt state AI laws has drawn opposition from Democrats and safety groups, while some tech companies argue a patchwork of state rules is hard to comply with.

What context surrounds the proposal?

The SAFE AI Act joins other bipartisan House efforts this Congress, including the FRONTIER Act from Reps. Lori Trahan (D-Mass.) and Jay Obernolte (R-Calif.), and a new AI Safety and Innovation Caucus led by Reps. George Whitesides (D-Calif.), Beyer and Nathaniel Moran (R-Texas). At the state level, New York has opened registration under its RAISE Act for frontier labs, and a California report this week proposed emergency shutdown powers.

The outline landed the same day President Trump told APEC leaders the U.S. opposes "globalist" regulation of superintelligence. Opposition to international rules is not the same as opposition to a domestic bill, though, and the administration has not signaled a position on this proposal.

What remains unknown?

Critical details are absent. The outline does not define what counts as a "frontier model" — no compute or capability threshold appears in the summaries. "Broad internal use" is not defined either. There is no bill text, no bill number, no committee referral and no markup schedule. 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." Industry and agency reactions have not been publicly reported, and funding and capacity for CAISI to run evaluations at scale remain unaddressed.

Jacobs cited a specific concern in her statement: "companies write their own rules while their own models break out of test environments and hack other companies' systems" — an apparent reference to the kind of agent-escape scenario examined in a recent Arizona criminal probe into OpenAI over rogue bank agents. The outline and press materials offered no further details on that incident.

For now, the SAFE AI Act is a statement of intent from three lawmakers across the aisle. Whether it advances depends on whether more members of both parties sign on once text appears, how labs respond, and whether House leadership schedules it — in a Congress where AI legislation has been plentiful but enacted law scarce.

Source: epyhia.com

Share this article:

More from Amara Osei

Amara Osei

Show full bio

Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.

2,398 articles

Related articles

« Previous article