AI Safety Board would regulate frontier models

Summary#

S. 5576, the Artificial Intelligence Risk Management and Security Act of 2026, would create a permanent advisory board inside the Department of Commerce and add duties for certain AI developers and critical-infrastructure operators. Its stated goal is to address risks from frontier AI models and AI agents. It has been introduced in the Senate and is not law; whether it passes is uncertain.

  • The Secretary of Commerce would establish the Artificial Intelligence Safety Board. NIST would supply its staff and support.
  • The Board would develop technical standards, including standards for evaluating AI risks and securing AI testing environments. The Secretary would adopt proposed or modified standards by rule, and developers would have to follow standards that apply to them.
  • A developer of a frontier AI model would have to give the Board access to the model and materials needed to run it at least 45 days before introducing it into commerce. The bill does not require the Board to approve a release.
  • Developers, as the bill defines that term, would have to publish, follow, and file a safety plan for each AI system or model they create, substantially modify, or use to train or evaluate other AI models.
  • Certain developers and operators of critical infrastructure would have to report confirmed AI safety or security incidents: within 30 days, or within 72 hours if the incident poses an imminent threat to national security, critical infrastructure, or public safety.
  • NIST would create a public database of selected incidents and AI flaws. The bill also directs NIST to develop guidance and a profile and documentation template for AI agents; these are not approval requirements.

What it means for you#

  • Frontier AI developers: They would have to provide pre-release access to covered models, follow applicable standards, and prepare and file safety plans. The term “developer” is limited in the bill to a developer of a frontier AI model, though the safety plans cover each system or model that developer creates, substantially modifies, or uses in training or evaluation.
  • Critical-infrastructure operators: Operators using AI with industrial control systems or other operational technology would have to report confirmed incidents within the stated deadlines.
  • Other organizations and researchers: Private and public organizations, civil society groups, and academic researchers could voluntarily share information about suspected or confirmed incidents and flaws.
  • AI agent users and organizations: NIST would develop a risk profile and a common documentation template. The bill does not require organizations to use them.
  • The public: The bill would create a database, but NIST would decide which reports or flaws belong in it. Published descriptions must generally protect the identities of affected parties.
  • Developers that violate applicable standards: The Secretary would enforce compliance. A developer could face a civil penalty of up to $250,000 per violation; each day a violation continues counts separately. The Attorney General could bring a civil case at the Secretary’s request.

Money#

The bill requires agency staff and support, but the available material gives no cost estimate.

  • NIST must provide staff and other support for the Board; the bill does not say how many staff or what that will cost.
  • The Secretary may use secure research computing environments on a reimbursable basis and make them available on a cost-recovery basis.
  • The bill sets civil penalties but does not estimate penalty revenue. It does not specify an appropriation or total federal spending.

What is unclear#

  • The bill does not set a precise test for which models count as “frontier” or meet its serious-risk threshold; the Board would develop evaluation processes.
  • The Secretary would decide the form and manner for filing safety plans. The bill does not spell out how much detail plans must contain or how confidential business information in them would be protected.
  • The bill does not describe how the Board would handle or secure the model materials it receives, beyond requirements related to classified information.
  • The bill does not set a detailed process for deciding whether a developer’s safety plan or compliance with a standard is adequate, or provide an appeal process.
  • It is unclear how “each violation” would be counted for penalties beyond the rule that each continuing day is a separate violation.

Case for#

  • A possible argument for the bill is that common technical standards could help developers assess risks and secure testing environments in more consistent ways.
  • Required incident reports from specified developers and infrastructure operators could give NIST information about risks that voluntary reports might miss.
  • A public database of selected incidents and recurring flaws could help others learn from problems, while the bill requires anonymization of affected parties.
  • Board access before a frontier model enters commerce could support safety and security testing before release.

Case against#

  • Developers would face mandatory standards, safety plans, model access, and reporting duties, but the bill leaves important details to later standards and agency decisions.
  • The Board would receive access to model weights and other materials before release. The bill does not specify detailed safeguards for these materials beyond its classified-information rules.
  • The Board’s work could involve sensitive or proprietary information, and the bill allows closed sessions on several grounds. That may limit public visibility into parts of its work.
  • The bill requires federal staff and new agency work without giving a cost estimate or specifying an appropriation.