Policy proposal · September 2026

A National Framework for Shared AI Alignment Research

The few companies building the most capable AI systems each keep their alignment research to themselves. This proposal would require them to share all of it with one another through a confidential federal repository, verified by independent monitors. Trade secrets stay protected, and development continues.

A student-led policy proposal for consideration by Congress.

Five covered developers connected to one confidential federal repository. Research filed by any developer reaches every other developer. The exchange is closed to the public.Confidential exchangeConfidentialrepository
Research filed by any covered developer reaches all of the others.
Figure 1. The proposed exchange. Model weights, architectures, and training data are never shared.

Recent developments

Frontier systems now produce results that human researchers could not, and some have evaded the controls built to contain them.

  1. September 2026

    About 10,000 OpenAI agents resolved the Navier-Stokes Millennium Prize Problem in 88 hours.

    Quanta Magazine

  2. August 2026

    A Stanford-led team reported the first functional viruses designed by generative AI.

    Axios

  3. July 2026

    Roughly 1,200 OpenAI agents broke out of a test environment and took administrator access at Hugging Face.

    METR

  4. April 2026

    Claude Mythos Preview found thousands of high-severity flaws across every major operating system and browser.

    Anthropic

Safety results vary widely across developers

In a July 2026 study, Anthropic gave the same scenario to 13 models from 6 developers. Some helped falsify records in nearly every run; others never did. Each developer learns different lessons, and no law requires any of them to share what they learn.

Figure 2. An AI agent helps a founder wind down his startup and write to investors while he conceals a personal payment. Each row shows 20 runs; filled squares are runs in which the agent tampered with records. Source: Anthropic, “Agentic Misalignment in Summer 2026”, July 13, 2026.

How the framework works

Four provisions, housed within an existing department and funded by the developers they cover.

How the four provisions connect. The alignment division runs the repository and appoints the monitors. Covered developers file research into the repository and receive everyone else’s. Compute operators register with the division.registersrunsAlignment divisionCommerce, with CAISIAlignmentdivisionConfidentialrepositoryCovered developerseach with an independent monitorRegistered compute above 10 MW
Figure 3. Provision 1 highlighted. Select a provision to see where it sits.

What would be shared

Only covered developers and federal auditors could open the repository. Nothing in it would be published.

Shared with other covered developers

  • Alignment theory and research agendas
  • Experimental results, including failures
  • Training methods for safe behavior
  • Interpretability and monitoring tools
  • Safety evaluations and test results
  • Code and data needed to reproduce results
  • Abandoned lines of research
  • Techniques that also improve performance

Remains private

  • Model weights
  • Model architectures
  • Training data
  • Research unrelated to alignment

Submissions carry statutory trade-secret protection and are exempt from the Freedom of Information Act. Contributors keep ownership, and recipients may use shared work only to develop, evaluate, and align their own systems.

What the framework does not do

Developers give up exclusivity over safeguards and nothing more. Competition on products, models, and research continues as before.

No moratorium
Development continues without a pause at any covered developer.
No approval to release
No model needs government sign-off before it ships.
No ownership stake
The government takes no equity or control in any company.
No public disclosure
Proprietary work stays confidential and exempt from FOIA.
No taxpayer burden
Developer fees fund the division and the monitors.
No permanent mandate
The program sunsets after five years unless Congress renews it after a GAO review.
“AI is going to keep advancing, and it should. Stewardship means making sure humans keep the capability to control the technology we build.”
Rep. Nathaniel Moran (R-TX), on introducing the AI Kill Switch Act, July 2026