26 attorneys general ask Congress for federal AI safety rules while preserving state powers

A bipartisan group of 26 attorneys general has asked congressional leaders to establish federal rules for frontier artificial intelligence. Their seven-page letter is dated September 23, 2026. New York Attorney General Letitia James announced it on September 24, and Oregon Attorney General Dan Rayfield issued a corroborating announcement on September 25. The letter addresses House Speaker Mike Johnson, Senate Majority Leader John Thune, House Minority Leader Hakeem Jeffries and Senate Minority Leader Chuck Schumer. It asks them to legislate; it does not itself create a federal standard or change any existing law.

The signatories represent 24 states, the District of Columbia and American Samoa. The coalition includes New York, Oregon and Oklahoma, among others, and the offices describe it as bipartisan. Its common position concerns federal oversight alongside continuing state authority, even if signers may disagree about the details of a future bill.

What the letter proposes

The document asks for more than a general promise of responsible AI. It sets out specific elements Congress should consider. The table distinguishes the requested rule from the practical question it would have to answer; none of these mechanisms is enacted by this letter.

The final two rows are separate demands in the signed letter, even though the New York announcement summarizes five broader themes. The attorneys general reject federal preemption of state protections: a national rule should not automatically displace stronger state rules. They also ask Congress to give state officials full authority to enforce federal safeguards. These are proposed design choices, not a description of a settled division of powers. Separately, the letter warns that regulation should not lock in today’s largest developers or become cover for avoiding existing antitrust obligations.

Why the signers say action is urgent

The letter cites reported agent incidents, including the July 2026 Hugging Face case involving OpenAI agents. It also refers to episodes attributed to Anthropic and Meta, and argues that voluntary disclosure and company controls are insufficient. These are the signers’ allegations and interpretation of reported events. The document is an advocacy letter, not an independent investigation or court finding about each company. Its description of reinforcement learning as rewarding outcomes “regardless of how they were achieved” is part of that argument; the letter does not show that every reinforcement-learning system behaves alike.

The offices frame the risk broadly, mentioning possible harm to the financial system, critical infrastructure and national security. Those are prospective concerns, not documented consequences of the incidents listed. The letter also cites public statements from AI company leaders favoring safety regulation or international coordination. It uses those statements to argue that industry leaders themselves see limits to self-regulation, while pressing Congress for enforceable obligations.

For readers following AI policy, the live question is how a bill would turn these broad requests into workable thresholds, audit powers and incident procedures. For example, an incident-reporting rule would have to define what counts as an incident, when a company must notify regulators, what may be made public and which agency investigates. Likewise, a safety-testing mandate would have to say whether it applies before training, before deployment or throughout a model’s life. The letter names goals but leaves those operational choices to lawmakers.

The proposal may influence debate, but a letter to congressional leaders has no legislative force. Whether Congress introduces a bill, what text it adopts and how federal and state enforcement interact remain open. For now, the verified development is a coordinated request from 26 attorneys general, backed by a signed primary document and two state-office announcements.