White House Finalizes Frontier AI Safety Framework, Then Says It Will Keep the Details Secret

Overview

A team from the Trump administration met this week with senior representatives from Anthropic, OpenAI, Google and Meta to review a draft framework under which AI firms would voluntarily submit frontier models to the government for safety testing before public release, according to SiliconANGLE. The review was hosted by the Office of the National Cyber Director, the outlet reported. Once finalized, however, the administration does not plan to make the framework public, according to Tech Policy Press, which reported that “the White House now says it will keep the framework a secret.”

What We Know

The framework traces to Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” which President Trump issued in June, according to Tech Policy Press. The order itself directs the government to “develop and maintain a classified benchmarking process to assess the advanced cyber capabilities of AI models” and lets developers “provide the Federal Government with access to covered frontier models” for “a period of up to 30 days before they plan to release such models to other trusted partners,” according to the White House. The order text itself states: “Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models,” according to the White House.

Tech Policy Press reported that a National Security Agency-led group within the government was instructed to build the framework, and that the process “operated informally, without published criteria, defined timelines, or any legal basis beyond the government applying pressure.” The outlet also reported that roughly 100 organizations currently have some form of access to reviewed models, but that “there are no published eligibility criteria” governing who is on that list.

The meeting this week, held on a Tuesday according to Fortune, brought administration officials together with representatives from OpenAI, Anthropic, Google, Meta, and Nvidia, Fortune reported. Fortune reported that the framework “targets closed-source models demonstrating state-of-the-art capabilities and presenting national security risks,” and that “open-weight models appear excluded from oversight requirements.”

The episode has real-world precedent. Tech Policy Press reported that the run-up to this framework included “a nineteen-day shutdown of Anthropic’s frontier models via an export control order” and “a two-week gated rollout of OpenAI’s GPT-5.6.” SiliconANGLE reported that Anthropic’s development of Mythos, a model “not released to the public due to its ability to unearth vulnerabilities in software,” first triggered the government’s concern, and that the administration later “implemented export controls on a derivative model known as Fable, the public version of Mythos, due to fears that foreign adversaries might try to use it to attack U.S. companies and infrastructure.” SiliconANGLE also reported that the White House separately “told OpenAI to stagger the release of its latest model, GPT-5.6.”

Reaction

The plan to keep the finalized framework confidential has drawn criticism from lawmakers and policy researchers. Fortune reported that the advocacy group Americans for Responsible Innovation said, “If only tech companies know what’s in the rulebook, it doesn’t work.” R Street’s Adam Thierer told Fortune the administration’s approach appears “far more arbitrary and burdensome” than prior policy. Representative Lori Trahan, a co-sponsor of the bipartisan FRONTIER Act, told Fortune that AI governance “belongs in a civilian agency, where it can be seen and questioned, not buried inside the national security apparatus.”

What We Don’t Know

Neither the White House nor the companies involved have published the finalized text of the framework, so the specific evaluation criteria, the threshold that would designate a model as “covered,” and the process for approving which organizations get early access to reviewed models all remain undisclosed. It is also not yet clear how the 30-day review period set out in the executive order will be enforced or extended in practice, since participation is voluntary and the order rules out a mandatory licensing regime.

Background

This is not the administration’s first attempt at a pre-release review structure for frontier models. A draft executive order that would have required up to a 90-day government review period was scrapped in May after direct lobbying of the president by Elon Musk, Mark Zuckerberg and David Sacks, as previously reported. Executive Order 14409, signed the following month, revives the concept with a shorter 30-day window. The framework also follows two ad hoc episodes it now aims to formalize: the temporary export-control shutdown of Anthropic’s Mythos-derived Fable model, and the staggered rollout of OpenAI’s GPT-5.6 to trusted partners before its public release.