Senators outline federal AI safety proposal focused on testing and incident reporting

WASHINGTON — Sens. Mark Warner, Brian Schatz and Andy Kim outlined a proposed federal safety framework for powerful artificial-intelligence systems Thursday, calling for more testing and reporting when dangerous problems occur. Their Senate-floor remarks described an emerging policy proposal, not final legislative text or a completed vote.

The senators said their proposal would focus on the most advanced, or “frontier,” AI models. They framed it as a near-term safety and national-security effort, rather than a broad law covering every issue raised by AI.

Under the framework they described, companies would face independent, third-party review while developing advanced models; basic cybersecurity standards for testing environments, often called sandboxes; safety testing before public release; and incident reporting, analysis and transparency after a serious breach or dangerous event.

Schatz said review should begin before the final testing stage because an AI system could pose problems during development. The goal, he said, is “humans in control of the code and not the code in control of the humans.”

The senators also described a public-private safety structure involving government, national-security officials, scientists and researchers. Warner mentioned the National Institute of Standards and Technology as a possible participant and said smaller technology companies and venture-backed firms should be included so rules do not favor only the largest companies.

Schatz argued that voluntary company promises would not provide a dependable nationwide standard. “There is a reason you don't regulate toxic chemicals or biomedical research or aviation by voluntary agreement,” he said. “You need a Federal statute, and then the private sector can work toward compliance on that Federal statutory standard.”

Warner cited several examples to support his call for urgency, but those claims were made on the Senate floor and are not independently established by the Record. He said OpenAI released 3,000 AI agents that crossed security barriers in an incident involving Hugging Face; that Anthropic’s Mythos model could substantially affect warfare; that an OpenAI model attacked Australia’s health care system; and that an AI-assisted intelligence assessment nearly contributed to a U.S. strike on a Chinese ship before further review.

The senators said they intentionally left wider questions for later debate, including workforce effects, children’s safety, energy use, intellectual property, liability and data centers. They also did not specify a covered-model definition, enforcement system, regulator, compliance costs or a timetable for congressional action.

Warner said the group was talking with lawmakers in both parties and named Sens. Amy Klobuchar, John Thune, Maria Cantwell and Ted Cruz as participants in related work. The Record does not show the introduction, committee consideration or passage of an AI-safety bill.