AI companies get classified threat intel—but public gets anonymized incidents only

What it does

This bill creates a federal framework for tracking and testing artificial intelligence systems before they are released to the public. It establishes an AI Risk Board to evaluate which AI models pose serious national security or public health risks, requires developers to submit frontier AI models to the National Security Agency for review 21 days before public release, creates a public database of AI security and safety incidents, and sets up a pilot program to share classified threat intelligence with AI companies about foreign adversaries targeting their systems.

Why we flagged it

The bill's core function is pre-release AI testing and incident tracking (regulatory), but a substantial portion establishes a classified intelligence-sharing pilot program that runs to private AI companies—a public-private intelligence partnership that is unusual and asymmetric in its disclosure protections.

What the text implies

• The bill exempts all threat intelligence shared with AI companies from FOIA disclosure (5 U.S.C. § 552(b)(3)), creating a permanent classified channel between NSA and private firms with no public audit trail of what foreign threats are being disclosed or to whom.
• Companies participating in the intelligence-sharing pilot gain liability protection (no cause of action) for sharing information with government or other companies, but the bill does not reciprocally require companies to disclose what they learn about foreign threats to the public or to non-participating competitors.
• The AI Risk Board includes both government officials and private AI developers/experts on the same body, with conflict-of-interest rules but no explicit prohibition on board members voting on standards that benefit their own companies' products.
• The 21-day pre-release NSA review window is mandatory but the bill does not specify what NSA can do with the model after review, whether it can be retained, copied, or used for intelligence purposes—only that access must be provided.
• The public incident database is anonymized by default and requires developer/provider consultation before publishing safety incidents, giving companies a de facto veto over disclosure of incidents involving their systems.

Section numbers refer to the bill text the analysis read — linked under Primary records below.

Who it affects

Citizens gain transparency through a public incident database and pre-release testing that may prevent harmful AI deployment. However, the bill grants broad classified-information access to private AI companies, exempts shared threat intelligence from public disclosure, and shields companies from liability for sharing information with government—creating asymmetric information flows that favor industry over public accountability.

Who stands to gain

• AI model developers and providers (frontier AI companies)
• Cybersecurity and infrastructure security contractors
• National security contractors with cleared personnel

Named in the bill

National Institute of Standards and Technology (NIST), National Security Agency (NSA), Cybersecurity and Infrastructure Security Agency (CISA), Artificial Intelligence Risk Board, Artificial Intelligence Security Center, Department of Commerce, Department of Treasury, Attorney General, Director of National Intelligence, Federal Bureau of Investigation

Where it stands

• Jul 21, 2026 — Introduced · Congress.gov: “Introduced in Senate”
• Jul 21, 2026 — Referred to Senate Committee on Commerce, Science, and Transportation · Congress.gov: “Read twice and referred to the Committee on Commerce, Science, and Transportation”

Dates and quoted wording are Congress.gov's action record; the timeline shows status changes, not every procedural step.