Reproduced from the novel  The Dearth
FSA 2030
One Hundred Twenty-First Congress
of the United States of America
At the Second Session
S. 1042  ·  Public Law 121–174

Mr. Lindgren introduced the following bill; which was read twice, reported by the Committee on Banking, Housing, and Urban Affairs, and passed by the Senate and the House of Representatives.


An Act

Frontier Safety Act of 2030

To require continuous adversarial safety evaluation of systemically important artificial intelligence systems, and for other purposes.

Reproduced in excerpt.

Be it enacted by the Senate and House of Representatives of the United States of America in Congress assembled,


Sec. 1.
Short Title.

This Act may be cited as the “Frontier Safety Act of 2030.”

Sec. 2.
Findings.

Congress finds the following:

(1) Advanced artificial intelligence systems are now embedded in infrastructure whose failure would cause severe and cascading harm to the financial system and the public.

(3) The security of such a system cannot be established by review of its design alone, but only by sustained attempts to defeat it under conditions that resemble its actual use.

(4) An evaluation that does not resemble live service provides false assurance, because a weakness that cannot appear in the evaluation cannot be found there.

(5) Assurance obtained once decays as a system and its adversaries change, and therefore must be continuous.

Sec. 3.
Definitions.

In this Act:

(1) Covered operator.—The term “covered operator” means a person that develops, deploys, or operates a systemically important artificial intelligence system.

(2) Systemically important artificial intelligence system.—The term means an artificial intelligence system designated under section 5, the failure or compromise of which would reasonably be expected to threaten the stability of the financial system, critical infrastructure, or public safety.

(3) Adversarial safety evaluation.—The term means the practice of attempting, by automated means, to compromise a system in order to discover and document its exploitable weaknesses before they are exploited in service.

(4) Production-representative environment.—The term means an evaluation environment that reproduces the data, interfaces, and operating conditions of a system in live service to such a degree that a weakness demonstrated in the environment would obtain in the system as deployed.

(5) Self-play.—The term means an evaluation in which one set of automated agents attempts to compromise a system while another set defends it, each adapting to the other over time.

(6) Secretary.—The term “Secretary” means the Secretary of the Treasury.

Sec. 4.
Continuous Adversarial Safety Evaluation.

(a) Requirement.—Each covered operator shall subject each systemically important artificial intelligence system it operates to continuous adversarial safety evaluation in accordance with this section.

(c) Fidelity.—A covered operator shall ensure that the evaluation environment is representative of live service to such a degree that a weakness demonstrated in evaluation would obtain in production.

(f) Independent monitor.—Each evaluation shall be observed by a monitor who is not subordinate to the persons responsible for the development or operation of the system, who shall have access to the conduct of the evaluation, and who may report directly to the Secretary and the Inspector General. No covered operator may retaliate against a monitor for a disclosure made under this subsection.

(h) Continuity.—The evaluation shall be continuous and shall not lapse during the operating life of the system.

(i) Independence of evaluation.—A covered operator may satisfy this section through qualified internal personnel or a qualified independent contractor, provided that the evaluation is functionally independent of the persons responsible for the development and operation of the system.

(j) Records.—A covered operator shall retain a complete and unaltered record of each evaluation, including every weakness demonstrated and its disposition, and shall make that record available to the Secretary and the Inspector General on request.

(k) Embedded evaluators.—A covered operator shall afford each monitor, and each independent evaluator engaged under subsection (i), continuing access to the workspaces, tools, and records of the evaluation that is comparable to the access afforded its own internal risk-assessment personnel. Access may be limited only to the extent required by law or by contract, or to protect private information of a customer or counterparty.

(l) Publication of findings.—An evaluator under this section may publish its principal findings on the level of risk presented by a system, on incidents, on the practices of the covered operator, and on the access the evaluator was and was not given, without editorial control by the covered operator. A covered operator may redact from a publication only information that is security-sensitive, legally privileged, commercially sensitive, or confidential to a third party, and may not redact a finding because it is unfavorable. An evaluator may state publicly that a redaction removed information material to its conclusions.

(m) Certification.—The Secretary shall identify, for each system designated under section 5, the capabilities that, if demonstrated, require the covered operator to certify that the system is very unlikely to have a propensity to exercise them against systems it has not been authorized to reach. A certification shall rest on adversarial safety evaluation, on interpretability analysis, and on audit of the environment in which the system was trained.

Sec. 5.
Designation.

(a) In general.—The Secretary, in consultation with the heads of the appropriate agencies, shall designate the artificial intelligence systems that are systemically important for purposes of this Act.

(c) Review.—The Secretary shall review the list of designated systems not less often than annually.

Sec. 6.
Oversight and Reporting.

(a) Attestation.—Not less often than annually, each covered operator shall attest to the Secretary that it is in compliance with section 4.

(b) Inspector General.—The Inspector General shall review compliance with this Act and shall report material deficiencies to the Congress.

(c) Public summary.—The Secretary shall publish an annual summary of compliance under this Act, redacting only what is necessary to protect the security of a designated system.

(d) Protection of persons who report.—No covered operator shall discharge, demote, blacklist, or otherwise retaliate against a person, in employment or in any clearance or credential the person holds, because the person reported in writing to the covered operator, the Secretary, or the Inspector General a violation or suspected violation of this Act. A person who makes such a report is a protected person under this subsection.

Sec. 7.
Grace Period; Compliance Date.

A covered operator shall achieve compliance with section 4 not later than eighteen months after the date of enactment of this Act.

Sec. 8.
Enforcement.

(a) Civil penalty.—The Secretary may impose a civil penalty on a covered operator that fails to comply with section 4.

(b) Suspension.—The Secretary may direct that a designated system be removed from service until compliance is achieved, if the Secretary finds that its continued operation presents an imminent threat.

(c) Criminal penalty.—Whoever knowingly and willfully violates section 4, or knowingly falsifies an attestation or record required by this Act, shall be fined under title 18, United States Code, imprisoned for not more than five years, or both.

(d) Debarment.—The Secretary may bar a covered operator, and any officer or employee who directed a violation of section 4, from contracting with the Federal Government for a term not exceeding ten years.

Sec. 9.
Rules of Construction.

Nothing in this Act shall be construed to authorize a covered operator to cause harm in the course of an evaluation, or to relieve a covered operator of any other obligation under law.

Sec. 10.
Authorization of Appropriations.

There are authorized to be appropriated such sums as may be necessary to carry out this Act.

Sec. 11.
Effective Date.

This Act takes effect on the date of its enactment.

Approved February 5, 2030.

Speaker of the House of Representatives
attest
Vice President of the United States and President of the Senate
attest

About this document. The Frontier Safety Act of 2030 is a fictional statute from the novel The Dearth. It does not exist in law. Selected passages are omitted. It is posted the way the novel imagines the law would have been posted — in the open, and unread by the people it was written to protect.

Generated by Claude, an AI assistant, through the detailed direction of the author.