AI Threat Output and Monitoring Incident Containment Act
- Sponsor
- Rep. Maloy, Celeste [R-UT-2]
- Committees
- Science, Space, and Technology Committee (primary)
- Last action
- Jul 27, 2026
Bottom line
This bill mandates a new federal program to proactively test and regulate advanced AI models developed by large entities to prevent AI-related nuclear security risks, imposing penalties for non-compliance while protecting proprietary information.
What it actually does
This bill directs the National Nuclear Security Administration, through the Secretary of Energy and the National Laboratories, to establish the Advanced Artificial Intelligence Nuclear Evaluation Program (AAINEP). The program is designed to test and evaluate advanced artificial intelligence models for their potential to cause "AI nuclear incidents," including through "jailbreaking" techniques and "loss-of-control scenarios." It mandates participation from "large advanced artificial intelligence developers," outlines prohibitions and enforcement mechanisms for non-compliance, and establishes confidentiality protections for submitted information. The Secretary is also required to provide Congress with legislative recommendations based on program findings.
Proponents argue
Supporters would argue this bill is a critical, proactive step to address the existential risks posed by advanced AI, particularly concerning national security and nuclear safety. They would emphasize the need for government oversight and rigorous testing, including red-teaming, to ensure AI models cannot be misused or become uncontrollable in ways that could lead to catastrophic "AI nuclear incidents."
Opponents contend
Opponents might contend that the bill imposes overly burdensome regulations on AI developers, potentially stifling innovation and competitiveness in the rapidly evolving AI sector. They might also raise concerns about the scope of the definitions for "advanced artificial intelligence" and "AI nuclear incident," the potential for government overreach, and the confidentiality provisions, suggesting that mandatory access to model weights could expose proprietary information.
The bill is concise and clearly structured, making it relatively accessible for an informed reader to understand its core provisions and implications within a reasonable timeframe.
Section 2(a)
Establishment of the Advanced Artificial Intelligence Nuclear Evaluation Program (AAINEP)
This provision directs the Secretary of Energy, acting through the National Laboratories, to establish the Advanced Artificial Intelligence Nuclear Evaluation Program (AAINEP) within 90 days of the bill's enactment. The program's core functions include testing and evaluating advanced AI for the likelihood of "AI nuclear incidents," implementing protocols for "jailbreaking" techniques, facilitating independent third-party assessments, providing reports to participants on testing outcomes, and developing recommendations for containment and mitigation strategies for AI nuclear incidents and loss-of-control scenarios.
Supporters argue
Supporters argue this is a necessary and timely measure to leverage federal expertise in nuclear security to proactively address emerging AI risks, ensuring that advanced AI systems are rigorously tested before deployment.
Critics contend
Critics might argue that establishing such a program within 90 days is an aggressive timeline for a complex technical undertaking, potentially leading to rushed implementation or insufficient resources.
Tradeoffs
Balances the urgency of addressing AI risks with the practical challenges of establishing a new, highly technical program within a federal agency.
Section 2(b) and (c)
Mandatory Participation and Prohibited Actions for Large AI Developers
This provision mandates that "large advanced artificial intelligence developers" participate in the AAINEP. It also explicitly prohibits these developers from failing to participate, making false or fraudulent statements, concealing material information, impeding the program, failing to provide secure access to artificial intelligence (including model weights and versions without safety mitigations), or altering/destroying relevant records. These prohibitions are designed to ensure full cooperation and transparency from regulated entities.
Section 2(d)
Enforcement Mechanisms for Non-Compliance
This section grants the Secretary of Energy robust authority to enforce the program's requirements. It allows the Secretary to assess civil penalties of up to $1 million for each violation, with continuing violations (like failure to participate) counting as a separate offense each day. The Secretary can also issue subpoenas to compel the production of AI models, information, and documentation necessary for testing if a developer fails to comply. Additionally, the Secretary may refer matters to the Attorney General to initiate civil actions in federal court for violations.
Section 2(e)
Confidentiality Protections for Submitted Information
This provision establishes that information submitted to the AAINEP by participants is exempt from disclosure under the Freedom of Information Act (FOIA). However, it outlines several exceptions where disclosure is permitted: if required by court order or legal process, with the participant's consent, upon request from a Congressional committee, or if the Secretary determines it's necessary for national security, public safety, or critical infrastructure. Anonymized or aggregated information that does not include trade secrets can also be disclosed. Participants must be notified if their information is to be disclosed under certain exceptions.
Section 2(f)
Legislative Recommendations to Congress
This provision directs the Secretary of Energy to assess data collected from the AAINEP within one year of enactment to identify trends, capabilities, and risks associated with artificial intelligence, including "evaluation awareness" and the potential for "AI nuclear incidents." Based on this assessment, the Secretary must develop and submit recommendations to Congress for future legislation aimed at preventing such incidents. These recommendations may cover areas such as appropriations, certification/licensing processes, the establishment of new federal agencies or offices, or broader regulatory oversight. The Secretary is also required to provide annual updates to these recommendations until the program terminates.
Section 2(h)
Program Termination Date
This provision sets a definitive end date for the Advanced Artificial Intelligence Nuclear Evaluation Program (AAINEP). The program is mandated to terminate seven years after the date on which the Act is enacted into law. This establishes a sunset clause for the entire initiative.
Section 2(i)
Key Definitions for the Program
This subsection provides critical definitions that determine the scope and applicability of the entire Act. Key terms defined include "advanced artificial intelligence" (AI trained with computing power greater than 10^26 operations or as specified by the Secretary), "AI nuclear incident" (a broad range of scenarios involving AI and nuclear weapons/materials, Restricted Data, loss-of-control in nuclear systems, or foreign adversary interference), and "large advanced artificial intelligence developer" (entities developing/modifying advanced AI for commerce, with over $2 billion in AI investments over five years). Other definitions include "loss-of-control scenario," "jailbreaking," "red-teaming," and "scheming behavior."
Defines "advanced artificial intelligence" as AI trained utilizing a quantity of computing power greater than 10^26 integer or floating-point operations, or other AI specified by the Secretary by rule. The Secretary's determination must consider developments in AI research, the need to include AI posing serious risk of an AI nuclear incident, and data from the Program.
Section 2(i)(1)
Why it matters:While standard legislative practice to place definitions at the end, the technical nature of the threshold makes it difficult for a non-expert to immediately grasp its regulatory implications.
Case for: Supporters would argue this provides a clear, objective, and quantifiable threshold for identifying the most powerful AI systems, while also granting the Secretary flexibility to adapt the definition as AI technology evolves.
Case against: Critics might argue that a fixed computational threshold, even with Secretarial discretion, could quickly become outdated or arbitrary, potentially missing dangerous AI models that achieve high capabilities with less compute, or over-regulating less risky models.
Estimated impact: This definition is foundational, determining which AI systems fall under the mandatory testing and evaluation regime, thereby shaping the entire scope of the program and its regulatory impact on the AI industry.
Defines a "large advanced artificial intelligence developer" as an entity that develops, owns, or substantially modifies advanced AI for utilization in interstate or foreign commerce, and has, collectively with its affiliates, expended not less than $2,000,000,000 on AI investments (including computing power, data procurement, personnel, infrastructure, and R&D) in the five calendar years immediately preceding the calendar year at issue.
Section 2(i)(13)
Why it matters:Standard practice for definitions, but the financial threshold is a key determinant of who is regulated, making it a critical detail often found in the fine print.
Case for: Proponents would argue this focuses the regulatory burden on the largest entities with the most resources and the highest potential to develop impactful AI, avoiding stifling smaller startups and concentrating oversight where risks are highest.
Case against: Opponents might contend that the $2 billion investment threshold is arbitrary and could be gamed, or that it might miss significant risks posed by smaller, highly innovative teams or state-sponsored actors not meeting the commercial criteria.
Estimated impact: This definition directly determines which companies are subject to the mandatory participation and enforcement provisions of the Act, thus shaping the competitive landscape and regulatory burden within the AI industry.
Broadly defines an "AI nuclear incident" to include an incident involving artificial intelligence and any of the following (or an attempt thereof): unlawful development/acquisition/utilization of a nuclear weapon or material; generation of Restricted Data; a loss-of-control scenario involving nuclear systems/stockpile; foreign terrorist organization or foreign adversary obtaining unauthorized access to, manipulating, or interfering with such a system; weaponization by such organization/adversary; or "scheming behavior" relating to nuclear assets.
Section 2(i)(3)
Why it matters:Standard legislative practice, but the comprehensive and somewhat technical nature of the definition makes it critical to understanding the bill's scope and the specific threats it aims to address.
Case for: Supporters would argue this comprehensive definition ensures that the program addresses a wide spectrum of potential AI-related nuclear risks, from direct weaponization to subtle "scheming behavior" that could undermine security.
Case against: Critics might argue that some aspects, like "scheming behavior" or "loss-of-control scenario" (especially with its $100M damage threshold), are vague or overly broad, potentially leading to subjective interpretations or over-reporting of incidents.
Estimated impact: This definition sets the boundaries for what the AAINEP is designed to prevent and detect, directly influencing the testing protocols, the types of risks the program prioritizes, and the reporting requirements for developers.