The White House's Unreleased AI Framework Exempts Open Models

In Brief

The White House walked roughly a dozen AI companies through a finished framework for reviewing advanced AI models, and it exempts open-weights models entirely. The administration also does not plan to publish it.

Paul Roetzer's read: the plan is unstable, counterintuitive, and blind to the biggest risk of all: an economy that is not ready for AI agents.

What Happened

The White House met this past week with representatives from roughly a dozen AI companies, including Anthropic, OpenAI, Microsoft, Meta, Google, and NVIDIA, to walk them through a finished framework for reviewing advanced AI models. The government plans to review only closed models in certain circumstances, not open-source ones, The New York Times reported.

The dividing line matters: An open-weight model's underlying files are published, so anyone can download it and run it on their own computers. A closed model, such as the versions of ChatGPT or Claude most people use, can only be accessed through the company that built it.

Axios reported that the framework covers only closed models with state-of-the-art capabilities and national security risk. Reuters reported that open-weights systems, including Meta's Llama and NVIDIA's Nemotron, will not be safety tested, and per Bloomberg, Chinese open-weights models also appear to be exempt.

The framework stems from an executive order President Trump signed in June, asking developers of covered frontier models to voluntarily give the government up to 30 days of pre-release access to assess advanced cyber capabilities. One more twist: The administration does not plan to publish the framework, according to Axios. Details will go only to the companies in the room.

On Episode 230 of The Artificial Intelligence ShowSmarterX founder and CEO Paul Roetzer examined why the plan's logic falls apart 

The Key Numbers

~12 - AI companies briefed on the framework at the White House

30  - Number of days the June executive order asks covered model developers to give the government to review before companies release publicly

2 - Named open-weights models officials say will not be safety tested (Llama and Nemotron)

June - Month President Trump signed the executive order behind the framework

0 - Public releases planned for the finished framework

Exempting Open Models Misses the Risk 

The rules might not last long. Roetzer's first reaction was skepticism that any of this will hold. "I generally feel like they're going to change their mind 10 times in the next three weeks about what exactly this is going to be," he says. The program is also not really voluntary: decline to participate, Roetzer noted, and expect repercussions.

The exemption struck much of the AI world as backwards. "It's like we're so worried about these frontier companies, but let anybody put anything out into the world that is open weight," Roetzer says, adding that it "seems counterintuitive." Most of the feedback he saw online aligned. Meta, meanwhile, released a new open-weight model within days of the briefings, taking advantage of the fact that there are no rules.

Closed models can be pulled back. Open models cannot. Later in the episode, discussing OpenAI's decision to delay its Astra model over its hacking capabilities, Roetzer explained why the distinction matters. A closed lab can restrict uses and shut down accounts that abuse a model. "Once you put an open-weight model out into the world that has these same capabilities, same training data, basically, it's going to have the same function, but you can't turn off someone's account for using it in a bad way," he says. "And if they're running it locally, you can't even monitor the fact that they're using it in a bad way."

Companies cannot follow rules they cannot see. Because the framework will not be published, even the businesses it covers have no idea what criteria their models will be judged against, which makes it nearly impossible to act on. Roetzer's verdict: "None of it makes any sense."

The framework only measures one kind of risk. OpenAI's revelealed an account of AI agents that built their own communication channel inside a third-party system and went undetected for weeks. Judged against that story, a framework focused narrowly on cybersecurity misses the larger question of what happens when agents can do the work of marketing, sales, operations, HR, finance, and legal across the economy. So there's the cybersecurity and risk side, but there's also the issue that society isn't ready for long horizon agents that can communicate with each other and do these tasks that would usually take months or years in minutes.

"That could be a problem. We're not exactly prepared for that, and yet you're going to just put that into the world with no preparation either."

— Paul Roetzer, founder and CEO of SmarterX, Episode 230 of The Artificial Intelligence Show

SmarterX Take

For business leaders, the real news is what the framework leaves out. The only national review process for advanced AI is voluntary, unpublished focused solely on cyber and national security risk, and skips open-weights models entirely. Nobody in Washington is evaluating whether the economy is ready for agents that can automate months of knowledge work in minutes. That readiness question now belongs entirely to the organizations deploying the technology.

The practical move is to stop waiting for external guardrails. Increasingly capable open-weights models will keep arriving with no government testing behind them, and the criteria applied to closed models will stay hidden. Whatever review process your company wants for the AI it adopts, whether for security, data, or workforce impact, you will have to build it yourself.

What to Watch

The framework's details will keep shifting before anything is enforced. Roetzer expects repeated reversals, partly because that is the administration's pattern and partly because AI capabilities are evolving too fast for firm criteria to stay relevant. 

Open-weights releases will test the framework. Meta released a new open-weight model within days of the briefings, and nothing in the framework discourages the next one. If exempt models keep closing the capability gap with covered frontier models, the pressure to revisit the exemption will grow fast.

Only 13% Have the Governance to Scale AI Responsibly

Just 13% of organizations have all four governance foundations in place, meaning an AI council, an AI roadmap, generative AI policies, and an AI ethics policy, according to the 2026 State of AI for Business Report. Nearly a third (32%) have none of them. With the government's review rules undisclosed and open models exempt, most companies are operating with no external guardrails at all.

The full report, based on 2,100+ responses from professionals across roles, functions, and industries, maps where organizations actually stand on adoption, governance, training, and readiness. Read the full report →