Trump may be forced to reveal secret rules feds use for AI safety testing

Ashley Belanger reports the nonprofit Protect Democracy sued four federal agencies to force release of the secret framework the Trump administration uses to safety-review frontier models before release. The group says almost no details have gone to the public or Congress: not the process, not which companies helped build it, not the legal authority, not how 'trusted partners' were chosen. It wants records by September 30 — the framework text, participation terms, participant identities, and how access to frontier models is granted or withheld — plus an injunction against withholding unclassified records.

The complaint names the Office of the National Cyber Director, OSTP, Treasury, and Commerce. Protect Democracy argues the White House is choosing which companies can release products and which customers get access, with no oversight, and that the process could favor firms Trump likes or punish others — pointing to a judge's ruling that blacklisting Anthropic was illegal. OpenAI is described as having a private agreement limiting distribution of cutting-edge models to government-vetted partners. The White House finished a voluntary pre-release review framework on August 3 and is using it alongside GOLD EAGLE, a July clearinghouse for flagging cyber vulnerabilities. A spokesperson said unclassified does not mean it will be broadcast. CISA 2015 liability protections that GOLD EAGLE reportedly depends on lapse unless Congress extends them, with votes possible September 30 or by December 11.

The lawsuit is a FOIA sledgehammer aimed at a voluntary regime that is acting like a licensing board. 'Unclassified but unpublished' is a governance choice, not a classification one. The spokesperson's line — we will not broadcast it — is the policy. Protect Democracy is asking a court to treat that as unlawful withholding.

GOLD EAGLE plus a completed August 3 framework, already in use, means the secrecy is not about a draft. It is about operating rules for who gets to ship. OpenAI's gated-partner arrangement, as alleged, is the concrete version: the state as channel partner for the most capable weights. Combined with the Anthropic blacklist episode, the conflict-of-interest story writes itself even if the framework is technically competent.

Signal Desk's read: the CISA clock is the leverage. If GOLD EAGLE needs those liability shields, Congress is being asked to extend a program whose membership and terms it has not seen. That is the cleanest oversight failure in the piece. California's SB 813, which Becker says was built in public, is the foil — not because Sacramento is wiser, but because a public standard can be argued with.

September 30 is both the records demand and a possible CISA date. Expect the White House to produce little, slowly, and to call the rest operational detail. The question is whether a judge treats 'voluntary' as a reason to hide the rules, or as a reason they cannot be secret.

After the government flagged Anthropic's Mythos 5 as too dangerous to release, the administration rushed a voluntary review process through the Center for AI Standards and Innovation, including tests with safeguards reduced. Hugging Face getting hacked by an OpenAI model is the recent example Protect Democracy uses to say blind trust is not available.