TheChek News Flow

A key agency has circulated a draft of the framework to OpenAI, Anthropic and Google

The Trump administration is close to finalizing its voluntary framework for AI companies to submit their most advanced models to the government before releasing them to the public, according to three people familiar with the matter.

The White House’s Office of the National Cyber Director circulated a draft of the framework to OpenAI, Anthropic and Google around two weeks ago, said the people, adding that the three companies jointly submitted their own edits. An executive order on AI signed by President Donald Trump in early June set a deadline of Aug. 1 for the creation of the framework, which aims to establish a more formal process for powerful new forms of AI capable of swiftly discovering cybervulnerabilities.

The framework would represent a more concrete set of rules for AI companies after weeks of uncertainty over how the Trump administration will manage the release of frontier models. The administration created the executive order largely in response to Anthropic’s frontier model, Mythos, which the company held back from wide public release because of its offensive cybersecurity capabilities. Since the signing of the order, the White House has imposed export controls on Anthropic over apparent security vulnerabilities in its most advanced models and has ordered OpenAI to do a staggered release of its latest model, GPT-5.6.

Those actions sparked criticism that the White House was imposing an ad hoc licensing regime on the AI industry. In contrast, the goal of the framework is to create a more standardized process for determining what constitutes a frontier model and allowing the government to review those models before sharing them with outside partners.

With days to go until the deadline, some AI companies have expressed concerns in meetings with White House agencies over how the framework will define frontier models, as well as whether it will treat closed- and open-source models differently, according to two people familiar with the matter.

There is also a looming question of how voluntary the framework will be in practice, especially after the White House’s decision to impose export controls on Anthropic. The Trump administration has maintained that adhering to it will not be mandatory. Still, the framework’s completion would represent a major win for the White House as it seeks to boost the U.S. AI industry amid rapidly accelerating competition from China.

The early June executive order was the result of weeks of debate between different White House agencies, as well as engagement with AI companies, over how to deal with frontier models like Anthropic’s Mythos. The Trump administration sought to maintain the hands-off approach to AI regulation championed by former AI and crypto czar David Sacks while still ensuring that it could supervise the new model’s national security risks.

Trump held off on signing the initial version of the executive order after a last-minute intervention from Sacks over concerns that it would be too onerous for companies. Later, Trump signed an updated version stating that AI labs would provide the federal government with access to “covered frontier models” for a period of up to 30 days before releasing them to other partners. The order, signed on June 2, provides a 60-day window for various government agencies to design the framework that would define frontier models, along with a government process for reviewing them.

Since June, the White House has held feedback sessions with companies on the framework, including a joint meeting on June 9, as it raced toward the Aug. 1 deadline. Still, some smaller AI companies worried that the framework would keep only the three largest labs—OpenAI, Anthropic and Google—in mind, according to two people familiar with the concerns, particularly for the definition of a covered frontier model.

There are different metrics to determine what to consider a frontier model, including the model’s size, how it’s deployed and its use-case–specific capabilities as measured by third-party evaluations. But if the definition focuses only on the leading, closed-source models such as Mythos and GPT-5.6, it could exclude open-source models that are often trained with fewer parameters or deployed on premise, rather than through cloud providers, even if they have comparable capabilities.

For AI companies, there could be advantages to inclusion in the program. The executive order directs more funding to supporting critical infrastructure and cybersecurity defenses, though it isn’t clear how meaningful those resources will be or how they will be shared with companies participating in the program.

Open-source models that the framework does not label as frontier models could be cut off from participation. Meanwhile, there has been a parallel discussion between companies and the government about whether open-source models should have exemptions from the voluntary review process based on their capabilities, to reduce the additional regulatory burden, The Information previously reported. Publicly available U.S. open-source options still lag behind their Chinese counterparts, and figures like Sacks argue for less regulation to ensure they can catch up.

Key details of how the model review process will work couldn’t be learned. The White House has communicated to some companies that the entity conducting the reviews will be a combination of the National Security Agency and the Center for AI Standards and Innovation, a small agency within the Commerce Department that has faced political pressure under the Trump administration, according to two people familiar.

Still, there is growing optimism that the White House understands CAISI’s technical expertise and importance, said one of the people.

Perhaps the most important unanswered question about the new framework: What will happen to companies that decline to participate in the voluntary model review process?

In a May press briefing ahead of the initial executive order’s release, a senior White House official didn’t directly answer a question about what would happen if a company did not submit to the guidelines. “It’s built to be flexible and adaptable and move at the speed of new technology development,” they replied. “We believe that this strikes the appropriate balance between security and innovation.”