The AI Threat Output and Monitoring Incident Containment Act, cited by its sponsors as the ATOMIC Act, is a bill in the 119th Congress that would direct the National Nuclear Security Administration, through the Secretary of Energy, to establish an Advanced Artificial Intelligence Nuclear Evaluation Program to test advanced AI systems for nuclear-related risks. It was introduced in the House on July 27, 2026 (Source: congress.gov).
Status and timeline
Rep. Celeste Maloy (R-UT-2) is the sponsor of record. The bill was introduced on July 27, 2026 and referred the same day to the House Committee on Science, Space, and Technology. As of August 1, 2026 the Congressional Research Service summary on Congress.gov is listed as in progress, and no Congressional Budget Office cost estimate has been posted (Source: congress.gov).
Rep. Sara Jacobs (D-CA-51) and Maloy announced the bill jointly on July 29, 2026, describing it as bipartisan and Jacobs as a co-lead (Source: sarajacobs.house.gov).
The full bill text has not been posted to Congress.gov. The provision-level statements below therefore rest on the sponsors' descriptions of the bill and on contemporaneous congressional reporting rather than on the enacted language.
Scope and covered developers
The sponsors state that the bill applies only to the largest developers of advanced AI systems, defined as those that have invested at least $2 billion in AI development over the previous five years. Covered developers would be required to provide the secure access and technical information necessary for evaluation, and the bill is described as including protections for proprietary business information and trade secrets (Source: sarajacobs.house.gov).
A cumulative-investment threshold of this kind differs from the training-compute triggers used elsewhere — California SB 53 defines a frontier developer at 10^26 FLOPs and adds a $500 million annual-revenue test only for its strictest tier, and EU AI Act (Regulation 2024/1689) uses a compute threshold for general-purpose models presenting systemic risk. See Compute Thresholds.
Key provisions
The Advanced Artificial Intelligence Nuclear Evaluation Program would use the Department of Energy's National Laboratories to test advanced AI systems for four categories of vulnerability named by the sponsors: generating sensitive nuclear information, bypassing safety guardrails, exhibiting deceptive behavior, and acting unpredictably in high-stakes environments. The program would also develop containment protocols, mitigation strategies, and technical standards addressing AI-related nuclear threats (Source: sarajacobs.house.gov).
Reporting by Nextgov/FCW describes participation as mandatory for large AI developers, with failure to participate or obstruction carrying a civil penalty of up to $1 million, and with each day a covered developer does not participate counted as a separate violation (Source: nextgov.com).
Reporting requirements
The Department of Energy would deliver an initial report to Congress within one year identifying trends, capabilities, and risks observed through the evaluation program, together with recommendations on whether additional legislation is warranted. Annual updates would follow for the program's stated seven-year duration (Source: sarajacobs.house.gov).
Reactions
Jacobs framed the bill in terms of pre-emptive identification of risk, saying "No one wants another Three-Mile Island or any other major nuclear incident" and describing the provisions as "common-sense guardrails" and "an important first step." Maloy emphasized existing federal capacity, saying that "the United States already has the world's leading nuclear security experts in our National Laboratories" and that the bill "builds on that expertise" so that Congress has "the information we need to make informed decisions about future AI policy" (Source: sarajacobs.house.gov).
The sponsors state that the bill has been endorsed by the AI Policy Network (Source: sarajacobs.house.gov). No opposing statements have been recorded in the available sources.
Related
The ATOMIC Act sits alongside other federal proposals that would place a government body between frontier developers and deployment decisions rather than relying on developer self-assessment, including AI Kill Switch Act (Lieu–Moran). Its use of national-laboratory testing capacity rather than a new evaluation body distinguishes it from proposals surveyed in Independent Verification Organizations (IVOs).
Open questions
- Whether the enacted text defines "advanced AI system" by capability, by training compute, or solely by the developer-investment threshold is not established by the sponsors' descriptions.
- How the program's evaluation findings would interact with existing classification rules governing nuclear weapons design information has not been described publicly.
- Whether the $2 billion investment threshold is measured on a cumulative or annual basis is not stated in the available sources.
Relationships
- instance-of: Frontier AI Governance
- related: Department of Energy (DOE) (would administer the evaluation program), OpenAI, Anthropic, Google DeepMind (developers likely to meet the covered-developer threshold), Dual-Use Frontier AI, Compute Thresholds, Independent Verification Organizations (IVOs), AI Kill Switch Act (Lieu–Moran), California SB 53, AI Incident Reporting