Biosecurity procurement requirements for covered artificial intelligence models

(a) Biosecurity evaluation framework (1) In general The Secretary of Defense shall develop a framework for the implementation of biosecurity evaluation standards and best practices relating to covered artificial intelligence models to mitigate biosecurity risks to the Department of Defense from the use of such models, including any risk presented by a malicious actor gaining unauthorized access to such models. (2) Coverage of relevant aspects of biosecurity risk The framework developed under paragraph (1) shall cover all aspects of biosecurity risks posed by and to covered artificial intelligence models, including the following: (A) Evaluation of the potential for misuse, including the capability of covered artificial intelligence models to materially enable, meaningfully facilitate, or significantly lower barriers to the design, acquisition, production, or dissemination of biological agents, toxins, or other biological threats. (B) Evaluation of model behaviors and failure modes relevant to biosecurity, including the propensity to provide actionable assistance for harmful biological activities, including through prompt-based elicitation, fine-tuning, tool use, or other methods of adaptation. (C) Biosecurity red-teaming and testing requirements, including the use of qualified subject-matter experts and test protocols designed to simulate realistic misuse attempts. (D) Mitigation measures and safeguards, including the deployment of technical and operational controls to reduce biosecurity risks, such as access controls, monitoring, logging, usage restrictions, capability gating, safe completion techniques, and limitations on high-risk functionality. (E) Risks relating to the unintended exposure, theft, or release of covered artificial intelligence models (including model weights, parameters, system prompts, fine-tuning data, or other sensitive model artifacts) and the biosecurity implications of such unauthorized access. (F) Supply chain and third-party risks relevant to biosecurity, including the risks posed by integration of covered artificial intelligence models into other systems and the risk that downstream deployments or failure by contractors to implement required biosecurity controls. (G) Processes for re-evaluation over time, including triggers for reassessment based on material changes to a model (including capability improvements, new tooling, fine-tuning, or deployment changes) and updated threat information. (3) Risk-based scaling The Secretary of Defense shall ensure that the framework developed under paragraph (1)— (A) prioritizes the covered artificial intelligence models that present the greatest potential biosecurity risk, based on threat reporting, risk assessments, and model capability evaluations; (B) establishes security and evaluation levels that scale to the biosecurity risk posed by a covered artificial intelligence model, taking into account the sensitivity of the mission context and the potential consequences of misuse; and (C) is designed to be practicable for acquisition and implementation, including through the use of tiered requirements and appropriate tailoring to the type of technology and deployment. (4) Consultation In developing the framework under paragraph (1), the Secretary of Defense may consult with— (A) such officials as the Secretary of Defense considers appropriate; and (B) any relevant industry, academic, and civil society experts with demonstrated biosecurity expertise. (b) Condition of procurement; implementation through acquisition regulations (1) Condition of procurement Beginning on the date that is 18 months after the date on which the Secretary of Defense issues implementing guidance under paragraph (2), the Department of Defense may not procure, obtain, renew, or extend the use of a covered artificial intelligence model (including through a contract for cloud-hosted model access) unless the covered artificial intelligence model has undergone biosecurity testing and evaluation consistent with the framework developed under subsection (a) and the Secretary determines that appropriate mitigation measures and safeguards have been implemented commensurate with the risk. (2) Implementing guidance and DFARS Not later than 180 days after the date of the enactment of this Act, the Secretary of Defense shall issue implementing guidance to carry out the requirements of this section. The Secretary may amend the Defense Federal Acquisition Regulation Supplement to incorporate requirements consistent with such guidance, including requirements applicable to contractors that develop, deploy, store, host, fine-tune, integrate, or provide access to covered artificial intelligence models for the Department of Defense. (3) Contractor assurances and documentation The implementing guidance under paragraph (2) shall include contractor assurance and documentation requirements sufficient for the Department to verify compliance, which may include— (A) summaries of testing protocols and results, including biosecurity red-teaming results, limitations, and remediation actions; (B) descriptions of mitigation measures implemented and evidence of their operation; (C) supply chain and access-control information relevant to the risk of unauthorized access to the model; and (D) such additional information as the Secretary determines necessary to evaluate biosecurity risk and compliance, consistent with protection of classified information and proprietary business information. (4) Waiver The Secretary of Defense may waive the requirements of paragraph (1) with respect to a specific procurement or use if the Secretary determines that exigent circumstances or overriding national security needs require such waiver. Any waiver under this paragraph shall be— (A) in writing; (B) time-limited; (C) accompanied by a description of compensating controls to manage biosecurity risk to the maximum extent practicable; and (D) provided to the congressional defense committees not later than 30 days after the waiver is issued. (c) Report Not later than 180 days after the date of the enactment of this Act, and annually thereafter for 3 years, the Secretary of Defense shall submit to the congressional defense committees a report describing— (1) the framework developed under subsection (a) and the status of implementation under subsection (b); (2) the categories of covered artificial intelligence models subject to the framework and any updates to the scope of coverage; (3) the extent to which the Department has incorporated biosecurity evaluation requirements into procurement processes and contract instruments; (4) challenges and resource needs for implementing biosecurity evaluations and mitigations; and (5) any recommendations for additional authorities or legislative actions. (d) Definitions In this section: (1) Artificial intelligence; machine learning The terms artificial intelligence and machine learning have the meanings given such terms in section 5001 of the National Artificial Intelligence Initiative Act of 2020 (15 U.S.C. 9401). (2) Biosecurity evaluation The term biosecurity evaluation means testing, assessment, and analysis conducted to identify and characterize biosecurity-relevant capabilities, misuse potential, and failure modes of a covered artificial intelligence model, including through red-teaming, and the evaluation of mitigation measures to reduce such risks. (3) Covered artificial intelligence model The term covered artificial intelligence model means an artificial intelligence model, or class of artificial intelligence models, designated by the Secretary of Defense for purposes of this section based on an assessment that the model’s biological or life-sciences-related capabilities, scale, access patterns, or deployment context could present a material biosecurity risk to the Department of Defense if misused or accessed by unauthorized persons.