H.R. 10981: American AI Security Act

This bill would require the Department of Commerce, working with the Department of Defense and several other federal agencies, to write rules for a special category of advanced AI systems called covered foundation models. These are broad-purpose AI models that meet certain size and capability thresholds, especially if they could create serious cybersecurity risks or help people develop chemical, biological, or radiological weapons.

What the government would have to do

Within 6 months of enactment, Commerce would have to issue regulations that:

• Define which AI foundation models are covered, including setting a minimum parameter threshold.
• Identify the types of capabilities that make a model especially risky, such as helping with cyberattacks or weapons development.
• Define what counts as a “substantial modification” of a covered model.
• Require developers to give the National Security Agency full access to the model and related information before initial deployment and before deploying a substantially modified version.

What AI developers would have to provide

Developers of covered models would need to provide information on:

• How the model was trained.
• What safety tests were performed and the results.
• How the developer plans to deploy the model.
• Who will be allowed to access the model.

Review before deployment

The bill would generally prohibit deployment of a covered model unless one of two things happens:

• The National Security Agency completes its review, provides recommendations, and the developer carries out those recommendations; or
• The agency fails to complete the review within the required time period.

After receiving access, the NSA would have 30 days to evaluate whether the model could:

• Help carry out or support a cybersecurity threat.
• Help find or exploit vulnerabilities in information systems.
• Help someone evade or bypass cybersecurity defenses.
• Provide information about making chemical, biological, or radiological weapons.
• Help plan or carry out the release of such weapons.

The NSA would then issue an unclassified report with a summary of the findings and recommendations, and it could include a classified annex. The evaluation could be extended by 30 days if needed.

Enforcement and appeals

The Secretary of Commerce could impose civil penalties of $500,000 per day for failing to provide required access or for violating the deployment restrictions.

The bill would also create a new Article I court, the Artificial Intelligence Security Review Court, to hear appeals from developers about the NSA’s recommendations. That court would review whether the recommendations were properly made and limited to the risks the law covers, but it would not review the agency’s classified methods and tools. The bill sets deadlines for filing briefs and for the court to issue decisions.

Other provisions

The bill would require Commerce to review and update the regulations after 2 years and then annually. It would also require a report to Congress within 18 months after the rules are issued, covering the regulation, how effective the evaluations have been, and any suggested legislative changes.

The bill would preempt state and local laws that impose transparency, disclosure, or reporting requirements on developers of covered foundation models regarding safety protocols, safety and security frameworks, capability evaluations, risk assessments, or rules issued under this section.

Relevant Companies

• None found