The Pulse
Schatz and Warner Seek 45-Day AI Checks Before Release
U.S. senators Brian Schatz and Mark Warner introduced legislation requiring frontier AI developers to give a proposed federal board access to models 45 days before release. The bill would replace the White House’s voluntary review framework
AI.info Team ·
Schatz and Warner challenge the White House’s voluntary approach
On September 24, 2026, U.S. Senators Brian Schatz and Mark Warner proposed a federal requirement for frontier AI developers to provide their models for government review before release. Their bill would require access at least 45 calendar days before a covered model enters commerce, a marked departure from the White House’s June policy, which calls for a voluntary review framework and says it does not authorize mandatory preclearance.
The proposal, the Artificial Intelligence Risk Management and Security Act of 2026, would establish an Artificial Intelligence Safety Board within the Department of Commerce. It was introduced in the Senate by Warner for himself and Schatz, and the bill text says it was read twice and referred to committee. The senators said they planned to speak on the Senate floor to press for its passage.
The board would receive model weights and operating files
The 45-day requirement goes beyond a briefing or summary: developers would have to make the model available to the board, including its weights, configuration files, runtimes and necessary software libraries. The bill defines a frontier model by its capacity to perform tasks posing serious risks to national security, economic security or public health and safety.
The proposed board would include representatives selected by NIST, Commerce, the Cybersecurity and Infrastructure Security Agency, the National Security Agency and the Treasury Department, as well as technical experts from outside government. It would develop technical evaluation and security standards, including protections for testing environments where a model might identify and exploit software vulnerabilities without direct human prompting.
Enforceable standards, but no stated release veto
Developers would also have to create and publish a Model Safety Plan for each AI system or model they create or substantially modify. The plans would include a risk assessment, mitigation measures and the identity of a corporate officer responsible for carrying them out. Under the bill, violations of applicable standards could draw civil penalties of up to $250,000 per violation, with each continuing day counted separately.
The text requires advance access and compliance with standards, but it does not expressly grant the board authority to approve or block a release. That distinction matters: the bill proposes mandatory testing access and enforceable rules, not an explicit government licensing decision on each model.
The June order draws a different line
President Donald Trump’s June 2 executive order directs agencies to design a voluntary framework under which developers could provide access to designated models for up to 30 days before sharing them with other trusted partners. The order also explicitly rules out using that framework to create a mandatory licensing, preclearance or permitting requirement for new AI models.
Schatz framed the bill as a response to security concerns around increasingly capable systems. “Our bill is about making sure humans remain in control of AI models and preventing future breaches with rigorous standards, testing, and oversight,” he said in the announcement. Warner said the bill would make companies’ most capable models subject to testing before public release, pointing to cybersecurity risks and security failures as reasons for federal rules.
The next step is congressional action: the senators are asking the Senate to pass the measure, while the legislation remains a proposal referred to committee. Its central dispute is now explicit in the documents: whether pre-release government access to the most powerful models should remain voluntary or become a legal obligation.