Cantwell Unveils AI Framework With NIST Standards and Continuous Model Testing

Senate Commerce Committee ranking member Maria Cantwell, D-Wash., released a six-point framework for regulating artificial intelligence on Wednesday that would give the federal government a major role in setting safety standards and testing frontier models for catastrophic risk, Roll Call reported. The framework does not include legislative text.

Standards, testing and audits

Under the plan, the National Institute of Standards and Technology would set standards for frontier models, including ways to address catastrophic risks such as loss of control and AI agents “escaping secure testing environments.” The framework calls for proportionate safeguards, with heightened review for systems that show dangerous capabilities or can improve themselves in ways that may bypass safety controls. The standards would also cover open-source models and be updated as risks evolve.

Covered models would face “continuous” testing by the government and third-party groups, with independent audits to confirm compliance. A Democratic committee aide said national security experts might be best placed to test whether a model can help design a biological or nuclear weapon, while third-party testers might handle questions such as whether agents can deceive their operators and escape their test environments.

Disclosure and enforcement

Frontier developers would have to publicly disclose potential risks and mitigations and report safety incidents, including “unsafe recursive self-improvement,” in which a model develops itself or other models without human instruction. The aide said enforcement would be worked out as the framework becomes legislation and could include civil or criminal penalties depending on the violation. The framework also covers defensive AI for critical infrastructure, children’s use of AI products, worker training, AI in consequential decisions and the U.S. role in international standards.

Why it matters

The approach contrasts with the White House and some congressional Republicans, who have favored voluntary standards and self-regulation, as in the voluntary self-policing accord signed with AI leaders. Senate Majority Leader John Thune has said he is working on an AI bill, but Cantwell is not part of those negotiations, and Commerce Chair Ted Cruz has said a federal standard would need to address the patchwork of state rules, such as California’s new AI safety executive order. The framework could set the tone if Democrats win the Senate in the midterms and Cantwell returns to the chair. Earlier this year she co-sponsored a bill from Sen. Todd Young, R-Ind., to codify NIST’s AI standards center.

What to watch

Whether the framework becomes one bill or several, whether any Republicans sign on, and how its testing and incident-reporting ideas compare with whatever Thune’s bill proposes.