Skip to content
Menu
Menu

House Bill Would Require NSA Review Of Powerful AI And Override State Disclosure Laws

The NSA would get up to 90 days to test new models before release, and the bill would override state AI safety disclosure laws.

Reps. Josh Gottheimer (D-N.J.) and Mike Lawler (R-N.Y.) filed the American AI Security Act, a bill that would bar AI frontier developers from releasing the most powerful AI models until the National Security Agency (NSA) has tested them and the developers have made the changes the NSA recommends. The bill was introduced in the House and sent to committees, and it has not been voted on.

If the NSA misses its deadlines, a developer could release a model without waiting for the review.

The bill leaves it to the Commerce Department to determine which models count as the most powerful. Commerce would write those rules within six months of the bill becoming law, working with the Defense Department, the NSA, and five health, defense, energy, and AI standards agencies.

The rules would set a minimum model size, measured by the number of internal settings that training sets. They would also identify any ability that could pose a serious national security risk. Finally, they would define what counts as a major model update. Each major update would need its own NSA test.

Under the bill, even giving early access to business customers counts as a release.

The bill would also stop states from requiring AI frontier developers to disclose their safety plans, test results, or risk assessments.

The NSA would test models before release

Under the bill, a developer would have to give the NSA full access to a model before release. The developer would also have to tell the NSA:

  • how the developer trained the model
  • what safety tests the developer ran, and the results
  • how the developer plans to release the model
  • who would get access to the model

The NSA would then test whether the model could do any of the following:

  • help carry out cyberattacks
  • find or exploit security flaws
  • get around security defenses
  • help someone build or use chemical, biological, or radiological weapons

The NSA would then recommend changes to the model to prevent those capabilities.

The NSA would have 30 days from getting access to test the model. It could extend that to 60 days by notifying the developer in writing within the first 30 days. It would then have 30 more days to send the developer a report with any recommended changes. In all, the NSA could take up to 90 days.

A developer could release the model only after making the recommended changes. If the NSA missed either deadline, the developer could release the model without implementing any changes.

Under the bill, the Commerce Department could fine a developer $500,000 for each day the developer failed to give the NSA access or the required information. Commerce could also fine a developer $500,000 for each day after the developer released a model before the NSA finished its review, or without making the changes the NSA recommended.

The bill would make the NSA review mandatory. Today, government review is voluntary. Under an executive order President Trump signed on June 2, developers of models the NSA flags as a serious cyber risk may give the government up to 30 days of access before public release. The government can’t block the release.

States could no longer require safety disclosures

The ban on state disclosure rules would reach cities and counties as well as states. It would also cover laws already in force.

Law firm Davis Polk reported in May on two state laws that require these kinds of disclosures. Under laws in California and New York, AI developers with more than $500 million in yearly revenue must publish their safety plans. Any developer of the largest AI models must also publish a report on each new model. The California law took effect in January 2026. The New York law takes effect in January 2027.

Developers could appeal to a new court

Under the bill, a developer could appeal the recommended changes within 30 days of receiving them. The bill would create the Artificial Intelligence Security Review Court to hear these appeals. The court would have five federal judges.

The court could decide only two questions. The first would be whether the NSA tested the model and sent its report the way the bill requires and on time. The second would be whether the recommended changes deal only with the national security risks the NSA tested. The court could not review how the NSA reached its recommendations, including any classified testing methods.

Clayton Rifkind

Clayton Rifkind is the Founder and Senior Editor of AI Risk Today. He also advises on business development for ESG Today, a leading source of ESG investment news and research for institutional investors and corporate leaders. He has 20+ years of experience in B2B technology, leading strategy and execution of go-to-market plans across software, enterprise platforms, and mobile applications. He founded two consultancies advising startups and Fortune 1000 companies, including Autodesk, Intel, and Microsoft. He began his career in the San Francisco advertising scene working with brands such as Hewlett-Packard, Intel, Microsoft, Symantec, and Wells Fargo. Clayton launched AI Risk Today in 2025 after two decades of watching enterprises adopt transformative technologies, and seeing how often risk, governance, and compliance considerations lagged behind. His reporting draws on primary sources including regulatory filings, court documents, and official announcements, with a focus on what AI developments mean for the executives accountable for managing them. Reach him at Reach him at [email protected] or on LinkedIn.

Essential AI Risk Intelligence

Daily insights on AI governance, regulation, and enterprise risk management. Trusted by Chief Risk Officers and compliance leaders globally.

By subscribing, you agree to receive our daily newsletter. Unsubscribe anytime.

Advertise with AI RIsk Today, Today!