Advanced AI Nuclear Evaluation Program

Summary#

This bill creates a new federal program run by the Department of Energy (through the National Laboratories) to test advanced artificial intelligence for risks of causing or enabling a nuclear-related incident. It requires large AI developers to give the program access to models and related materials, sets confidentiality rules, allows subpoenas and fines for non‑compliance, and directs the Department to produce recommendations for containment and for possible new laws. The program lasts up to seven years and must report to Congress within one year.

• Main change: Establishes the "Advanced Artificial Intelligence Nuclear Evaluation Program" to test and red-team advanced AI for risks tied to nuclear weapons, materials, facilities, or loss-of-control scenarios.
• Mandatory participation: Companies defined as "large advanced AI developers" must participate and provide access to models, including model weights and versions without safety mitigations when needed for testing.
• Enforcement tools: The Secretary of Energy may assess civil penalties (up to $1 million per violation), issue subpoenas to compel access, and refer civil enforcement to the Justice Department.
• Confidentiality: Data submitted is generally exempt from public disclosure but the Secretary may disclose information in specific cases (court order, participant consent, congressional request, or for national security/public safety), and may publish anonymized or aggregated data that does not include trade secrets.
• Reporting and policy work: The Secretary must assess program data, recommend legislation or regulatory changes within one year, update recommendations annually, and may issue policies or standards for National Labs based on program findings.

What it means for you#

• Large AI companies / developers

  • If you meet the bill’s definition of a "large advanced AI developer" (including at least $2 billion spent on AI development over five prior years), you must participate in the program and provide access to models and documentation needed for testing.
  • You may need to provide model weights and unmitigated versions of models for testing and red-teaming.
  • You face possible fines, subpoenas, and civil action if you refuse or obstruct participation.

• National Laboratories and the Department of Energy

  • Labs will run tests, red teams, and blind model evaluations to look for nuclear-related risks and for behaviors like "jailbreaking" or "scheming."
  • Labs will prepare participant reports that list test outcomes and the risks and safety measures tested.
  • The Department will develop containment and contingency recommendations and may issue departmental policies or standards for lab conduct.

• Congress and federal policymakers

  • The Department must submit a report within one year identifying trends and legislative recommendations (for example, licensing, oversight, or new agency creation), and update those recommendations annually.

• Public safety and national security

  • The program is designed to identify and reduce risks that advanced AI could generate technical information, facilitate weaponization, cause loss-of-control events, or enable foreign adversaries or terrorist organizations to access or misuse nuclear systems.

• Other developers, researchers, and the public

  • The bill focuses on large developers and National Labs. There is no direct new requirement for small AI firms or individual researchers in the bill text.

Expenses#

No publicly available information on a formal cost estimate or fiscal note is included in the bill text.

• The bill requires the Department of Energy and National Laboratories to run testing, red-teaming, reporting, and rulemaking work. This would likely require staff time, lab resources, computing resources, and possibly new contracts or facilities.
• Participating companies may incur costs to prepare and securely provide model weights, unmitigated model versions, and documentation.
• The bill allows civil penalties (up to $1 million per violation) but does not specify whether penalty revenue would offset program costs.
• Because the bill removes the Paperwork Reduction Act requirement for this program, administrative reporting may proceed without that clearance process.

Proponents' View#

• The bill appears intended to reduce the risk that advanced AI could enable or directly cause nuclear-related incidents by creating a formal testing and evaluation program.
• Supporters may argue that structured red-teaming, blind evaluations, and tests for jailbreaking and scheming behavior will reveal dangerous capabilities before models are widely deployed.
• The program’s requirement for developer participation and the Department’s authority to compel access via subpoena could help ensure tests use real models and realistic inputs.
• The bill’s reports and legislative recommendations aim to give Congress and agencies information needed to consider further laws or regulatory frameworks for high-risk AI.

Opponents' View#

• One concern is that the bill requires companies to share model weights and unmitigated versions, which may expose trade secrets or sensitive intellectual property; the bill offers limited protections and allows disclosure in several circumstances.
• The Secretary has broad discretion to disclose participant information for national security or public safety and to respond to congressional requests; the standards for those disclosures are not tightly specified in the bill.
• It is unclear how the program will treat classified data, proprietary systems, or foreign-based developers and whether legal or operational conflicts could arise.
• The bill leaves some key thresholds and definitions to future rulemaking (for example, the computing‑power threshold for "advanced AI" and other determinations), so exact scope and coverage are uncertain until the Secretary issues rules.
• The civil penalty levels and the overall enforcement structure may raise questions about whether they are adequate or whether legal challenges would follow if companies resist compelled disclosure.