AI Safety Legislation: Warner and Schatz Seek NSA Pre-Release Assessments for Frontier Models

Two Democratic senators are forcing a vote this week on a bill that would require frontier AI developers to submit their models for NSA assessment before release, a significant shift from the current optional testing regime. The Warner-Schatz proposal also establishes a federal AI safety incidents database and tasks NIST with building a risk-assessment framework, signaling that federal oversight of AI safety is moving from voluntary toward mandatory.

What the Warner-Schatz bill requires

The legislation, introduced as a Senate unanimous-consent measure by Sens. Mark Warner of Virginia and Brian Schatz of Hawaii, mandates that frontier AI firms such as OpenAI and Anthropic have the National Security Agency evaluate their models prior to release. This replaces the Trump administration’s voluntary testing approach, which the senators argue is insufficient given the pace of AI development. Recent hacking incidents and a high-profile researcher resignation at Anthropic have amplified concerns.

NSA pre-release assessments for frontier models

Under the bill, the NSA would assess frontier AI models before public deployment. Sen. Warner emphasized practical risks: “I’m less worried about some science-fiction scenario where AI ends humanity than I am about a very real scenario where a powerful AI system breaks into a bank, disrupts a water system, or compromises some other piece of critical infrastructure.” The assessment requirement directly targets companies building the most capable AI systems.

Federal safety database and NIST framework

Beyond NSA vetting, the bill establishes a federal database for AI safety incidents to track and disclose security events. It also directs the National Institute of Standards and Technology to develop an AI risk-assessment framework, and authorizes the Commerce Department to coordinate testing and reporting with foreign partners. These provisions create a structured accountability layer that currently does not exist.

The unanimous-consent strategy

Warner and Schatz are using unanimous consent, an expedited procedure that allows passage if every senator agrees. The tactic sets up a potential Senate confrontation ahead of a planned recess, particularly if some Republicans object to mandatory pre-release assessments. Sen. Chuck Grassley separately explored using the same mechanism for AI whistleblower protections.

Why AI builders should care

For developers and product teams at frontier AI labs, this bill represents the most concrete federal push yet to mandate pre-release security review. If enacted, companies would need to build NSA assessment timelines into their release planning and prepare to report safety incidents to a federal database. The risk framework from NIST could influence how safety claims are evaluated and audited.

Even if the bill does not pass this week, the Warner-Schatz effort signals growing bipartisan appetite for regulation. Warner has been in close contact with OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, and JPMorgan Chase CEO Jamie Dimon, suggesting business leaders are already engaging with these requirements.

Caveats to watch

The analysis relies on a single Washington Sun article describing the bill’s broad provisions. The full legislative text, enforcement mechanisms, and timeline remain unverified. The proposal faces an uncertain path: unanimous consent requires no opposition, and partisan dynamics could block passage. Builders should track the actual vote and any amendments that emerge.