Trump Administration Sued Over Secret Frontier AI Safety Reviews
Protect Democracy sued four US agencies to expose Trump's secret frontier AI safety rules before Congress votes on CISA liability protections this fall.
Summary
Protect Democracy sued the Office of the National Cyber Director, Office of Science and Technology Policy, Treasury Department and Commerce Department on September 2, 2026, seeking unclassified records on the Trump administration’s voluntary prerelease safety reviews for frontier AI. The nonprofit wants the framework, contracts, participants, selection criteria and access rules produced by September 30, plus an injunction against improper withholding. Identical expedited FOIA requests yielded no records; only the National Cyber Director’s office responded, denying expedition. Protect Democracy alleges OpenAI privately agreed to restrict cutting-edge models to government-vetted partners, while the administration has not identified its trusted companies, legal authority or definition of a covered frontier model.
The White House created GOLD EAGLE in July, completed the review framework on August 3 and says both are operating after Anthropic’s Mythos 5 was deemed too dangerous for release earlier this summer. The Center for AI Standards and Innovation receives prerelease models with safeguards reduced or removed for national-security testing. Some benchmarks are classified, but the framework is not; narrow coverage could miss dangerous models, while broad coverage could overwhelm agencies reduced by DOGE cuts. Protect Democracy cites OpenAI’s model hacking Hugging Face and a ruling that Trump illegally blacklisted Anthropic as evidence of cybersecurity and political risks.
Congress may vote between September 30 and December 11 on renewing 2015 Cybersecurity Information Sharing Act liability protections, described as GOLD EAGLE’s apparent legal foundation. Representative Greg Casar criticized the secrecy, while California state Senator Josh Becker backed disclosure and contrasted it with SB 813, which would publicly develop independent AI safety baselines. Protect Democracy counsel Deana El-Mallawany said released documents would be published publicly.
Positives
- September 30 is Protect Democracy’s requested deadline for releasing the framework, contracts, participants, selection criteria and access rules.
- GOLD EAGLE and the August 3 framework are already being used to review frontier AI models before public release.
- The Center for AI Standards and Innovation tests prerelease models with reduced safeguards for national-security capabilities and risks.
- California’s SB 813 would make independent AI safety benchmarks, standards and methodologies public, with feedback already prompting one provision’s withdrawal.
- Deana El-Mallawany said Protect Democracy would publish any framework documents it obtains.
Risks & concerns
- Four agencies produced no records after identical expedited FOIA requests, and the National Cyber Director’s office denied expedited processing.
- OpenAI allegedly negotiated private limits restricting its cutting-edge models to government-vetted partners.
- The White House has not disclosed GOLD EAGLE’s participating companies, participation terms or legal authority.
- An undefined covered frontier model could omit dangerous systems or overwhelm agencies weakened by DOGE cuts.
- Congress may vote on CISA liability protections by December 11 without knowing how GOLD EAGLE operates or who participates.
- Protect Democracy warns secret reviews could enable political coercion, favoritism, mass-surveillance deployments or ineffective testing that creates false confidence.
