Trahan publishes revised AI safety framework with penalties for near misses
The updated FRONTIER Act would give independent evaluators embedded access and federal officials enforcement powers. Full legislative text is still to come.
Loading page…
The updated FRONTIER Act would give independent evaluators embedded access and federal officials enforcement powers. Full legislative text is still to come.
Listen to this story
Rep. Lori Trahan’s revised FRONTIER Act framework, released November 19, 2026, would expand federal oversight of frontier AI developers after months of stakeholder feedback. It would treat near misses as reportable and punishable, empower federal officials to order risk corrections, and preserve stronger state safeguards alongside a federal minimum. The release is a section-by-section summary, not legislative text; Trahan’s office says the updated bill text will follow in the coming weeks, leaving details subject to change.
Reporting would cover models deceiving their developers, incidents affecting critical infrastructure, and events developers should have known about.
Automated AI research and development, including recursive self-improvement, would fall under catastrophic-risk safety, reporting, and assessment requirements.
Independent verifiers would receive embedded access and recommend corrective actions; government assignment mechanisms are intended to prevent a race to the bottom.
Rep. Lori Trahan released a revised FRONTIER Act framework on November 19, 2026, proposing penalties for AI near misses, stronger independent oversight and federal orders to correct safety failures. The release is a section-by-section summary, not the updated legislative text, which her office says will follow in the coming weeks.
The proposal would also preserve state laws that offer stronger protections, rather than make federal rules the ceiling for AI safety. States with weaker or no standards would be subject to the federal floor. That approach would let stronger state safeguards coexist with a national minimum for the developers covered by the legislation.
Trahan and Rep. Jay Obernolte introduced the original FRONTIER Act on July 23, 2026. That bipartisan bill set out tiered requirements for frontier developers, including risk-management frameworks, independent audits, incident reporting and ongoing assessments. The new summary revises that oversight approach after months of stakeholder feedback, according to Trahan’s office.
Near misses would trigger both reporting duties and penalties for frontier developers. Trahan’s office describes the penalties as an incentive to take precautions before a catastrophe occurs. The proposal therefore would not reserve accountability solely for incidents in which catastrophic harm has already materialized; a narrowly avoided failure could also carry consequences.
Reporting would reach beyond near misses to cover models deceiving their developers and incidents affecting critical infrastructure. A broader knowledge standard would also apply: developers would have to report incidents they “should have known” about, not only those they actually knew about. That wording would make awareness a wider obligation than acknowledging a known incident.
The updated framework separately identifies automated AI research and development, including recursive self-improvement, as a category of catastrophic risk. That category would be subject to the bill’s safety-framework requirements, reporting duties and independent assessments, rather than sit outside those obligations.
The American people are fed up with the biggest AI companies writing their own rules and grading their own homework. That approach failed on social media, and we can’t let it happen again with an even more powerful technology.
Rep. Lori Trahan, in her November 19 release
Independent verification organizations would receive guaranteed embedded access under the revised proposal. It also calls for mechanisms allowing the government to assign those organizations. Trahan’s office says the assignment provisions are intended to avoid a race to the bottom in verification, making the selection of outside evaluators part of the oversight design.
Those evaluators would also have a duty to recommend corrective actions to developers. Separately, when a frontier developer fails to adequately mitigate risk, the federal government could require specific corrective action. Civil penalties and court injunctions would back those orders, giving officials an enforcement route beyond requiring a company to disclose its safety practices.
The proposed standard for developers’ published safety frameworks is outcome-based: they must be reasonably designed to adequately mitigate catastrophic risk. Commerce would publish guidance on best practices, standards and methods for meeting that requirement. The framework thus pairs a risk-reduction obligation with government guidance on how companies could satisfy it.
To administer the proposed regime, the framework would formally establish an Office of AI Security within the U.S. Department of Commerce. A Senate-confirmed Assistant Secretary for AI Security would lead it, placing responsibility for the new oversight system in a dedicated office rather than leaving its administration unspecified.
The proposal also authorizes funding and special hiring powers for that office. Those provisions address the staffing and resources needed to run the regulatory system, alongside the new duties it would impose on frontier developers.
Loading discussion...
Join the conversation
Explain whether penalties would encourage caution or discourage disclosure.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.