Four federal agencies are facing a lawsuit that could force the Trump administration to disclose the hidden rules it says it uses to test frontier AI models before release. The complaint, filed by the nonprofit Protect Democracy, argues that the public, Congress, and even most of the AI industry have been left in the dark about who participates in the review process, what legal authority supports it, and how officials decide which models can be tested or approved.
Trump administration’s secret AI safety testing framework is under scrutiny
According to Protect Democracy, the administration has revealed “almost no details” about the framework, despite using it to evaluate advanced AI systems. The group says the process appears to involve a small set of “trusted partners,” but neither those participants nor the criteria for selecting them have been made public.
The lawsuit seeks an order requiring agencies to produce non-classified records by September 30. That includes the framework’s text, participation terms, the identities of participants, and the process used to grant or withhold access to frontier models. Protect Democracy is also asking a court to stop agencies from improperly withholding records that are not classified.
Why Protect Democracy says the secrecy matters
The group’s core argument is not simply that the government is being opaque, but that it is making consequential decisions about cutting-edge AI with little outside oversight. In a statement, Protect Democracy said the executive branch is effectively deciding which companies may release products and which customers can access technology that could affect American industry, national security, and economies abroad.
That creates a risk, the group says, that the process could be used unevenly or even corruptly. The complaint points to agencies including the Office of the National Cyber Director, the Office of Science and Technology Policy, the Treasury Department, and the Commerce Department as possible participants in a framework that may reward companies favored by Trump and slow those he dislikes.
GOLD EAGLE and the broader AI review effort
The lawsuit also targets the White House’s July launch of GOLD EAGLE, a clearinghouse that relies on industry partners to help agencies identify cybersecurity vulnerabilities across sectors. Protect Democracy says the administration did not identify which companies are involved, what terms they accepted, or what legal basis authorizes the program.
On August 3, the White House said it had completed the voluntary framework for reviewing AI models before public release. According to the source material, both the framework and GOLD EAGLE are already being used as frontier models continue advancing.
Some parts of the process are classified. The executive order described a “classified benchmarking process” meant to assess advanced capabilities and determine when models should be designated as frontier AI. But Protect Democracy says the framework itself is not classified, which makes the refusal to disclose it more difficult to justify.
What the complaint says is missing
- The text of the review framework
- The terms under which companies participate
- The identities of the participants
- The criteria for granting or denying access to models
- The legal authority used to run the program
The White House has resisted calls to share more, with a spokesperson saying that “just because things are unclassified, that doesn’t mean we are going to broadcast them to everyone.” Protect Democracy argues that this leaves the public unable to evaluate whether the system is effective, fair, or even legitimate.
Congress is being kept in the dark, too
The complaint says lawmakers are also blocked from assessing the administration’s AI safety framework. That is especially important, Protect Democracy argues, because Congress is now weighing whether to renew the Cybersecurity Information Sharing Act of 2015, which appears to be the legal basis allowing AI companies to share information with the government through GOLD EAGLE.
The group says the relevant liability protections have lapsed before and that Congress will soon have to decide whether to extend them again. The earliest vote could come on September 30, with the latest by December 11, depending on congressional scheduling.
Protect Democracy’s view is that lawmakers cannot make an informed decision about CISA renewal without knowing how GOLD EAGLE works, who is participating, and what authority the administration is invoking. Deana El-Mallawany, the group’s director of Impact Programs & Counsel, told Ars that neither Protect Democracy nor Congress has enough information to make a fully informed decision on extending CISA.
Politics, AI safety, and concerns over oversight
The lawsuit arrives amid broader political fights over AI safety and government influence over the industry. In a post on X last month, Rep. Greg Casar, D-Va., accused Trump of “completely failing to keep us safe from the dangers of AI” while taking “millions from AI billionaires.” He criticized what he called a “voluntary” review system that no one has seen.
California state Sen. Josh Becker also urged the court to grant Protect Democracy’s request. In a declaration, he pointed to SB 813, a California bill that would set up a process in which independent organizations establish AI safety baselines. Becker said California’s approach would make benchmarks, standards, and testing methodologies public, unlike the federal framework described in the complaint.
Becker also noted that public input has already changed the bill, with one provision withdrawn after backlash. In his words, “every step of SB 813’s development has been public,” and the state is accountable for the framework it has set.
What happens next
For now, the legal fight centers on whether the administration must disclose the non-classified side of its AI review system. Protect Democracy says it filed identical FOIA requests with four agencies, seeking only unclassified material and asking for expedited processing. So far, no agency has produced records, and only the National Cyber Director’s office responded by denying expedited treatment.
If the group wins access, it says it will publish the documents. Until then, it argues that Congress, researchers, businesses, and the public are left guessing about how one of the most powerful emerging technologies is being governed.
Source: Original report
Was this helpful?
Explore more: AI Automation Services More AI & Automation Tech News
Last Modified: September 3, 2026 at 1:51 am
6 views
