Section 1 Short title
This Act may be cited as the “AI Threat Output and Monitoring Incident Containment Act”.
AI Threat Output and Monitoring Incident Containment Act
A BILL
To direct the National Nuclear Security Administration through the Secretary of Energy to establish the Advanced Artificial Intelligence Nuclear Evaluation Program, and for other purposes.
Sec. 2 Advanced artificial intelligence nuclear evaluation program
In general— Not later than 90 days after the date of the enactment of this Act, the Secretary of Energy (in this section referred to as the “Secretary”), acting through the National Laboratories, shall establish a program, to be known as the “Advanced Artificial Intelligence Nuclear Evaluation Program” (in this section referred to as the “Program”), to carry out the following:
Test and evaluate artificial intelligence to assess the likelihood of an AI nuclear incident with respect to such artificial intelligence.
Implement in such testing and evaluation protocols to address practical jailbreaking techniques for such artificial intelligence, including such protocols for red-teaming by entities that have expertise the Secretary determines comparable to sophisticated malicious actors.
Facilitate to the extent practicable independent, third-party assessments and blind model evaluations of such artificial intelligence.
Provide to each participant in the Program a report that includes information relating to the following:
The outcomes from such testing and evaluation with respect to the artificial intelligence of such participant.
An identification of the risks and safety measures so tested and evaluated.
Develop for large advanced artificial intelligence developers and the National Laboratories recommendations on containment protocols, contingency planning, and mitigation strategies for AI nuclear incidents and loss-of-control scenarios, that are based on data from such testing and evaluation.
Based on the recommendations developed under paragraph (5), issue or revise such policies, orders, notices, manuals, guides, and technical standards under the Department Directives Program as the Secretary determines appropriate to govern the conduct of the National Laboratories with respect to AI nuclear incidents and loss-of-control scenarios.
Participation— Large advanced artificial intelligence developers shall participate in the Program.
Prohibitions— Large advanced artificial intelligence developers may not—
fail to participate in the Program as required by subsection (b);
knowingly make or cause to be made a materially false, fictitious, or fraudulent statement or representation with respect to such developer’s participation in the Program;
knowingly conceal from the Secretary or fail to disclose to the Secretary material information that the developer is required to provide in connection with the Program;
impede or obstruct, or attempt to impede or obstruct, the Program;
fail to provide secure access to artificial intelligence, including model weights and, as necessary for testing and evaluation under subsection (a), versions of such artificial intelligence without safety mitigations, information, or documentation necessary for such testing and evaluation; and
alter, destroy, falsify, or conceal a record relevant to the Program with the intent to impede or obstruct the Program.
Enforcement—
Civil penalty—
In general— If the Secretary determines after reasonable notice and opportunity for a hearing that a large advanced artificial intelligence developer has violated this section, the Secretary may assess on such developer a civil penalty of not more than $1 million for such violation. In the case of a continuing violation, including a failure to participate in accordance with subsection (b), each day during which the violation continues constitutes a separate violation.
Factors— In determining the amount of a civil penalty to be assessed under subparagraph (A), the Secretary shall consider the following factors:
The nature, circumstances, extent, gravity, and duration of the violation at issue.
The degree of culpability of the large advanced artificial intelligence developer at issue.
Previous violations, if any, of this section by such developer.
Good-faith efforts, if any, by such developer to comply with this section.
Whether such developer voluntarily disclosed to the Secretary such violation.
Any other factor that justice may require.
Subpoena— If a large advanced artificial intelligence developer fails to participate in accordance with subsection (b), or fails to provide access as required by subsection (c)(5), the Secretary may issue to such developer a subpoena to compel the production of such artificial intelligence (including model weights and related software), and of such information and documentation, as is necessary for testing and evaluation under subsection (a).
Civil action— If the Secretary believes that a violation of this section has occurred, is occurring, or is about to occur, the Secretary may refer the matter to the Attorney General to bring in an appropriate district court of the United States a civil action for relief.
Confidentiality—
In general— Subject to paragraph (2), information submitted through the Program by a participant in the Program is exempt from disclosure under section 552(b)(3) of title 5, United States Code.
Exceptions— The Secretary may not disclose information referred to in paragraph (1) unless any of the following requirements are satisfied:
Such information is required to be disclosed pursuant to a court order or other legal process.
The Program participant at issue has consented to such disclosure.
A committee of Congress has requested such information.
The Secretary determines such information should be disclosed to a third party or the public, for a purpose relating to the national security, public safety, or critical infrastructure of the United States or an ally of the United States.
Such information satisfies the following requirements:
Is anonymized or aggregated.
Does not include a trade secret or confidential business information of such participant.
Notice— If the Secretary is to disclose pursuant to subparagraph (A), (C), or (D) of paragraph (2) information referred to in paragraph (1), the Secretary shall as soon as practicable notify each Program participant whose information is to be so disclosed, of such disclosure.
Recommendations for legislation—
Report— Not later than 1 year after the date of the enactment of this Act, the Secretary shall carry out the following:
Assess data from the Program to identify trends, capabilities, and risks regarding artificial intelligence, including with respect to the following:
Evaluation awareness.
Potential for an AI nuclear incident.
Scheming behavior.
Develop based on such assessment recommendations for legislation to prevent such an incident, which may include legislation regarding any of the following:
Appropriations.
A certification or licensing process.
Establishing a new Federal agency, or establishing a new office in an existing Federal agency.
Regulatory oversight.
Submit to Congress a report that includes information relating to such recommendations.
Updates— Not later than 1 year after submitting the report under subparagraph (C) of paragraph (1) and annually thereafter until the Program terminates, the Secretary shall carry out the following:
Update based on data from the Program the recommendations referred to in subparagraph (B) of such paragraph.
Submit to Congress a report that includes information relating to such recommendations, as so updated.
Nonapplicability of the Paperwork Reduction Act— Subchapter I of chapter 35 of title 44, United States Code (commonly known as the “Paperwork Reduction Act”), does not apply to the Program.
Termination— The Program shall terminate on the date 7 years after the date of the enactment of this Act.
Definitions— In this section:
Advanced artificial intelligence—
In general— The term advanced artificial intelligence means any of the following:
Artificial intelligence trained utilizing a quantity of computing power greater than 1026 integer or floating-point operations (or another quantity of computing power specified by the Secretary under subparagraph (B)).
Other artificial intelligence specified by the Secretary under subparagraph (B).
Determinations— A determination of the Secretary under clause (i) or (ii) of subparagraph (A) shall be made by rule. In making such a determination, the Secretary shall consider the following:
Developments in artificial intelligence research, and in the testing and evaluation of artificial intelligence, occurring after the date of the enactment of this Act or after the date of the most recent determination under this subparagraph, as the case may be.
The need to include in the Program artificial intelligence that may pose a serious risk of causing an AI nuclear incident.
Data collected through the Program.
Affiliate— The term affiliate means an entity that owns or controls, is owned or controlled by, or is under common ownership or control with, another entity.
AI nuclear incident— The term AI nuclear incident means an incident that involves artificial intelligence and any of the following (or an attempt thereof):
The generation of technical information, instructions, or assistance that is reasonably likely to serve as a substantial factor in the unlawful development, acquisition, or utilization of a nuclear weapon or nuclear material (as such term is defined in section 831 of title 18, United States Code).
The generation of Restricted Data (as such term is defined in section 11 of the Atomic Energy Act of 1954 (42 U.S.C. 2014)).
A loss-of-control scenario that involves systems utilized in connection with such weapon, such material, a nuclear facility (as such term is defined in section 2332i of such title), or the nuclear stockpile.
A foreign terrorist organization or foreign adversary obtaining unauthorized access to, manipulating, or otherwise interfering with, such a system.
Weaponization of such weapon, material, facility, or stockpile, as the case may be, by such organization or adversary, as the case may be.
Scheming behavior relating to such weapon, material, facility, or stockpile, as the case may be.
Artificial intelligence— The term artificial intelligence has the meaning given such term in section 238(g) of the National Defense Authorization Act for Fiscal Year 2019 (Public Law 115–232, 10 U.S.C. note prec. 4061).
Computing power— The term computing power means the processing power and other electronic resources utilized to train, validate, deploy, and run artificial intelligence.
Critical infrastructure— The term critical infrastructure has the meaning given such term in section 1016 of the USA PATRIOT Act (42 U.S.C. 5195c).
Deploy— The term deploy means to release, sell, or otherwise provide access to artificial intelligence.
Evaluation awareness— The term evaluation awareness means the capability of artificial intelligence to determine such artificial intelligence is undergoing a test, evaluation, or assessment, and modify the behavior, output, or performance of such artificial intelligence in a manner that might cause such test, evaluation, or assessment, as the case may be, to yield a result that is not representative of the ordinary behavior, output, or performance of such artificial intelligence.
Foreign adversary— The term foreign adversary means a foreign government referred to in subsection (a) of section 791.4 of title 15, Code of Federal Regulations, or a successor regulation.
Foreign terrorist organization— The term foreign terrorist organization means an organization designated under section 219 of the Immigration and Nationality Act (8 U.S.C. 1189).
High-stakes context— The term high-stakes context means a context in which the operation or output of artificial intelligence could reasonably be foreseen to bear materially on any of the following:
The safety of human life.
The security, reliability, or integrity of any of the following:
Critical infrastructure.
A national security system (as such term is defined in section 3552 of title 44, United States Code).
Another system the failure or compromise of which would foreseeably endanger such life or cause substantial physical or economic harm.
Jailbreaking— The term jailbreaking means an adversarial technique to circumvent a safety restriction, alignment measure, or use restriction of artificial intelligence and elicit a harmful or prohibited behavior or output by such artificial intelligence.
Large advanced artificial intelligence developer— The term large advanced artificial intelligence developer means an entity that satisfies the following requirements:
Develops, owns, or substantially modifies advanced artificial intelligence for utilization in interstate or foreign commerce, including by initiating a training run of such artificial intelligence.
Had collectively with the affiliates, if any, of such entity expended in the five calendar years immediately preceding the calendar year at issue not less than $2,000,000,000 on investments to develop or modify artificial intelligence, including such expenses with respect to any combination of the following:
Computing power.
Data procurement.
Personnel.
Infrastructure.
Research and development, including algorithm development.
Loss-of-control scenario— The term loss-of-control scenario means a scenario in which artificial intelligence pursues an objective that is different from any of the objectives intended by a human developer or operator of such artificial intelligence, by—
behaving contrary to human instruction in a high-stakes context,
altering without authorization from such developer or operator, as the case may be, the constraints of such artificial intelligence, or
subverting an oversight or shutdown mechanism,
National Laboratory— The term National Laboratory has the meaning given such term in section 2 of the Energy Policy Act of 2005 (42 U.S.C. 15801).
Red-teaming— The term red-teaming means structured testing in which an entity utilizes an adversarial method to identify a flaw, vulnerability, undesirable behavior, or dangerous capability of artificial intelligence, including the potential, if any, for an AI nuclear incident involving such artificial intelligence.
Scheming behavior— The term scheming behavior means behavior that artificial intelligence carries out to deceive a human, including any of the following behaviors by such artificial intelligence:
Hiding a capability or objective of such artificial intelligence.
Attempting to subvert an oversight or shutdown mechanism relating to such artificial intelligence.
Substantially modify— The term substantially modify means, with respect to artificial intelligence, to carry out an action that costs not less than $5,000,000 and materially alters the functionality or performance of such artificial intelligence.