Kelly’s proposal puts federal examiners inside frontier AI companies
Sen. Mark Kelly’s three-pillar proposal would embed federal examiners in frontier AI companies, define safety standards and tighten liability, but it remains policy, not legislation.
By Amara Osei
3 min read

What's News
- Released Tuesday: Kelly’s framework has three pillars covering federal safety standards, on-site examiners and liability.
- Independent examiners would identify national security threats during model development rather than after deployment.
- The FRONTIER Act cited by Crowell & Moring has 11 bipartisan co-sponsors.
- No bill number, committee referral or co-sponsors were announced for Kelly’s framework.
- The proposal does not specify the employing agency, examiner powers, funding or a frontier-model threshold.
By Amara Osei
A three-pillar federal safety framework released Tuesday by Arizona Sen. Mark Kelly would place dedicated independent federal examiners inside companies developing the most powerful AI models. Titled A Safety and Security Framework for Frontier AI, the proposal marks a shift from company self-oversight to mandatory, on-site government oversight, Kelly’s press release said.
What does the framework require?
The framework sets three pillars: federal safety standards, embedded independent federal examiners, and liability. Punchbowl News reported the same three-part structure, while Kelly’s office said the approach mirrors how the government regulates nuclear materials and weapons systems.
Kelly said: “The rule has always been simple: the more dangerous it is, the more carefully you check it. AI should be no different.” The former Navy pilot and astronaut sits on the Senate Armed Services and Select Intelligence Committees.
Which gaps does Kelly’s plan target?
Kelly’s release identifies three gaps: no common metric for when a model becomes dangerous; company self-oversight with outside review only after the fact; and no standardized acquisition rules for military AI. The framework’s answer is a rulebook, a referee and accountability.
Under the proposal, standards would define required safeguards, while examiners would identify national security threats as models are developed rather than after deployment. New defense acquisition rules would require rigorous safety testing and hold contractors accountable for hiding or neglecting known risks.
What risks underpin the proposal?
Kelly’s release cites whistleblowers from top frontier AI companies who warned of unchecked acceleration of AI technology. It also cites “the Hugging Face attack where a powerful OpenAI model broke out of its controlled environment, exploited security vulnerabilities, and hacked into a separate AI company without being instructed to do so.”
The framework argues that companies now decide which models are dangerous and how much protection they need, often putting their bottom line ahead of public safety. Kelly’s release treats federal examiners as a check on those company decisions.
How does the plan fit Congress?
As new Democratic majorities prepare a wave of AI bills, Crowell & Moring’s client alert describes bipartisan interest in compulsory, risk-based frameworks for frontier models. The alert cites Rep. Jay Obernolte’s FRONTIER Act, which has 11 bipartisan co-sponsors, and Sen. Bernie Sanders’ Ban Artificial Superintelligence Act.
The alert also cites reports that Rep. Ro Khanna plans a Human Control Over AI Act and Rep. Ted Lieu plans to push for an AI Kill Switch Act. Kelly previously introduced the Make AI Work for Americans Act, which would tax leading AI companies to fund a federal worker trust fund.
His new framework goes further by proposing a direct federal regulatory presence inside frontier-model development environments. The proposal would move scrutiny from company-controlled review toward federal observers working alongside model builders.
What remains unresolved?
Kelly’s document remains a policy proposal, not legislative text, and has not been introduced as a bill. No bill number, committee referral or co-sponsors were announced, and the press release does not identify the employing agency, examiner powers, funding or the threshold that would define a frontier model.
Kelly’s office said the framework builds on his AI leadership and national security oversight experience. Its path forward depends on the legislative calendar, committee dynamics and whether the cited risks, including the Hugging Face incident, withstand independent scrutiny.
Congressional action would need to settle examiner powers, the employing agency and the model threshold before the proposal can operate as written. For frontier AI companies, it would put federal scrutiny inside development; until those details are settled, its near-term operating impact remains uncertain.
Source: epyhia.com
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,336 articles