Secure A.I. Development Act of 2026
- Sponsor
- Sen. Warner, Mark R. [D-VA]
- Committees
- Commerce, Science, and Transportation Committee (primary)
- Last action
- Jul 21, 2026
Bottom line
The bill aims to proactively mitigate national security and public safety risks posed by advanced AI by imposing pre-release testing requirements, enhancing vulnerability tracking, and fostering intelligence sharing between the government and AI developers.
What it actually does
This bill establishes new federal oversight mechanisms for advanced artificial intelligence (AI) models, focusing on security and safety risks. It creates an AI Risk Board, mandates pre-release access for frontier AI models to the National Security Agency (NSA), and establishes a public database for AI security and safety incidents. Additionally, it updates federal vulnerability management processes to better account for AI-related vulnerabilities and initiates a pilot program for intelligence sharing with private AI developers.
Proponents argue
Supporters argue that this bill is a critical step to ensure the responsible development and deployment of powerful AI systems, protecting national security, economic stability, and public safety from emerging threats. They contend that proactive measures, including government-industry collaboration and robust vulnerability management, are essential to address the unique and rapidly evolving risks associated with frontier AI models before they are widely adopted.
Opponents contend
Critics may contend that the bill's mandatory pre-release access requirements for the NSA could stifle innovation, create significant compliance burdens for AI developers, and raise concerns about intellectual property protection. They might also argue that broad government access to proprietary models could disadvantage U.S. companies in the global AI race or lead to over-regulation of a nascent industry.
The bill is moderately complex due to its technical nature and references to existing statutes, but an informed reader can understand its core provisions within a reasonable timeframe.
Section 3(b)
Establishment of Artificial Intelligence Risk Board
This provision establishes an Artificial Intelligence Risk Board within the National Institute of Standards and Technology (NIST). The Board will be composed of government officials and non-governmental experts, tasked with developing technical evaluations to identify high-risk AI capabilities and creating best practices for AI security and safety, including cybersecurity measures and risk evaluation processes for personnel.
Supporters argue
Proponents argue that establishing a dedicated expert board is crucial for proactively identifying and mitigating the complex and evolving risks associated with advanced AI, ensuring that technical evaluations and best practices are developed by a diverse group of stakeholders.
Critics contend
Opponents might argue that the Board's composition, particularly the inclusion of national security agencies, could lead to a bias towards security over innovation, or that its processes could become overly bureaucratic and slow to adapt to rapid AI advancements.
Tradeoffs
This provision balances the need for expert guidance on AI risks with concerns about potential bureaucratic inertia or a narrow focus if not managed effectively.
Section 3(c) and (e)
Mandatory Pre-Release Access for Frontier AI Models to NSA
This provision requires providers of 'frontier artificial intelligence models' to grant the National Security Agency's (NSA) Artificial Intelligence Security Center access to their models, including weights and configuration files, at least 21 days before public release. Failure to comply incurs a penalty of $100,000 per day, though providers are given a 7-day 'right to cure' by withdrawing the model and providing access.
Section 3(d)
Frontier Artificial Intelligence Model Registry
This provision mandates NIST to establish a public registry of frontier AI models. Providers of such models are required to register them before introducing them into interstate or foreign commerce. The registry will include rules for registration, contesting registration, removal, and attestation of submission to the NSA test-bed.
Section 4(a)
Publicly Accessible Database for AI Security and Safety Incidents
This provision directs NIST, in coordination with CISA, to establish mechanisms for voluntary sharing of information on confirmed or suspected AI security and safety incidents. This information will then be used to populate a publicly accessible database, ensuring anonymization of affected parties and exempting shared information from public disclosure to encourage participation.
Section 5(b) and (c)
Updating Vulnerability Management for AI Systems
This provision directs NIST to reform the National Vulnerability Database (NVD) to handle the increased volume of vulnerabilities identified by AI systems, including using AI to characterize vulnerabilities and incorporating AI security vulnerabilities. It also requires CISA to update the Common Vulnerabilities and Exposures (CVE) Program to identify and enumerate AI security vulnerabilities.
Section 7(c)-(i)
Pilot Program for Intelligence Sharing with AI Developers
This provision establishes a 3-year pilot program, led by the NSA, to assess the feasibility of securely sharing intelligence and threat information with 'covered persons' (non-federal U.S. persons developing AI for federal agencies). The program aims to help these entities secure their AI supply chains against malicious foreign activity, with strict rules on participation, intelligence handling, permissible use, and nondisclosure, while also protecting privacy and civil liberties.
Establishes a daily penalty of not less than $100,000 for each day a frontier artificial intelligence model is available through interstate and foreign commerce without having obtained the voluntary security guidance from the NSA test-bed, following a violation of the mandatory pre-release access requirement.
Section 3(e)(3)
Why it matters:This is standard legislative practice to place penalties within enforcement sections. However, the specific amount and daily nature of the fine are significant and could be easily overlooked.
Case for: Proponents would argue that a substantial daily penalty is necessary to ensure compliance with the mandatory pre-release access requirement, reflecting the high stakes involved in securing frontier AI models against national security threats.
Case against: Critics might argue that such a high daily fine could be disproportionate, particularly for smaller AI developers, and could create an overly punitive regulatory environment that discourages innovation or forces companies to operate outside the U.S.
Estimated impact: Potentially millions of dollars in fines for non-compliant companies, serving as a strong deterrent.
States that any information voluntarily shared with NIST for the AI security and safety incident database shall be exempt from public disclosure and withheld from the public, and shall not be deemed a waiver of any applicable privilege or protection, including trade secret protection.
Section 4(a)(3)(B)
Why it matters:This is standard legislative practice for confidentiality clauses within information-sharing frameworks. Its placement ensures it applies directly to the voluntary submission process.
Case for: Proponents would argue that this strong exemption from disclosure is absolutely essential to encourage private sector entities and researchers to voluntarily share sensitive incident information without fear of legal repercussions, competitive disadvantage, or public exposure.
Case against: No specific opposition is typically raised against this provision, as it is a necessary safeguard for voluntary data submission. However, some transparency advocates might generally prefer more public data, but would acknowledge the necessity of this clause for participation.
Estimated impact: Significantly increases the likelihood of robust participation in the voluntary incident database, leading to a more comprehensive understanding of AI risks.
Defines 'covered person' for the intelligence sharing pilot program as a non-Federal U.S. person who develops, deploys, or operates artificial intelligence models or critical enabling infrastructure AND provides these services to a Federal department or agency.
Section 7(a)(5)
Why it matters:This is standard legislative practice to define terms at the outset of a section. However, the specific criteria significantly narrow the scope of who can participate in the intelligence sharing program.
Case for: Proponents would argue that this narrow definition ensures that intelligence sharing is focused on entities directly relevant to federal government AI procurement and national security, minimizing risks associated with broader dissemination of sensitive information.
Case against: Critics might argue that limiting the pilot program to only those who provide services to federal agencies excludes a significant portion of the AI industry that develops frontier models, potentially missing opportunities to secure a wider range of critical AI systems.
Estimated impact: Restricts the intelligence sharing pilot program to a specific subset of the AI industry, focusing its resources and impact.