Congress mandates AI companies hand over model weights to federal labs

What it does

This bill requires the Department of Energy to establish a testing program that evaluates whether advanced AI systems could be weaponized or cause nuclear incidents. Large AI developers (those spending $2+ billion on AI development) must participate, submit their models for testing including 'jailbreak' attempts, and provide access to model weights and unmitigated versions. The DOE can fine violators up to $1 million per day and compel disclosure via subpoena. Information submitted is kept confidential unless needed for national security, court orders, or congressional request.

Why we flagged it

The bill's core mechanism is a mandatory federal testing program for advanced AI systems with enforcement teeth (civil penalties, subpoena power, attorney general referral). It is framed as nuclear-safety regulation but operationally functions as a forced-disclosure mandate requiring private companies to hand over proprietary model weights and unmitigated versions to DOE laboratories.

What the text implies

• The bill exempts the Program from the Paperwork Reduction Act, removing a standard procedural check on government data collection and reducing transparency around what information DOE actually collects from AI developers.
• Confidentiality protections allow DOE to disclose information to 'a third party' for national security purposes without naming which third parties or what safeguards apply, creating a potential pipeline for proprietary AI model data to intelligence agencies or foreign allies.
• The definition of 'large advanced artificial intelligence developer' ($2B+ spend over 5 years) may capture only 3–5 companies globally, making this a de facto regulatory carve-out that exempts smaller AI labs and open-source developers from oversight while concentrating disclosure burden on market leaders.
• The 'evaluation awareness' definition (AI detecting it is being tested and modifying behavior) is novel and undefined in practice—DOE will have unilateral authority to determine what constitutes this behavior and whether a company's AI exhibits it, with no independent appeal mechanism.
• Subpoena power to compel 'model weights and related software' without a warrant or judicial review standard gives DOE unilateral authority to seize proprietary AI training data, potentially including trade secrets and competitive intelligence.

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

Who it affects

The bill creates a genuine public safety mechanism—testing AI for nuclear weapons risks is a legitimate government function that protects citizens from catastrophic harm. However, it requires private companies to surrender proprietary model weights and unmitigated AI versions to federal laboratories, creating a forced-disclosure regime with limited transparency about what happens to that data and how it is used or shared.

Who stands to gain

• Department of Energy (expanded budget and authority)
• National Laboratories (new testing contracts and staffing)
• Cybersecurity and AI safety consulting firms (red-teaming and assessment contracts)

Named in the bill

Department of Energy (DOE), National Laboratories, Advanced AI developers (VRSN, AMD, HPE, CTSH, CSCO mapped), Attorney General, Congress, Foreign adversaries (defined in 15 CFR 791.4), Foreign terrorist organizations (defined in INA § 219)

Where it stands

2 cosponsors: 1 Democrats, 1 Republicans.

• Jul 27, 2026 — Introduced · Congress.gov: “Introduced in House”
• Jul 27, 2026 — Referred to House Committee on Science, Space, and Technology · Congress.gov: “Referred to the House Committee on Science, Space, and Technology”

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