On Tuesday, staffers from OpenAI, Anthropic, Google, and other tech companies convened in Washington to review the administration's new framework for reviewing advanced AI models. Now, the rest of the industry eagerly awaits answers on what will be included in the framework — a structured way for President Donald Trump's administration to evaluate AI models' cyber capabilities.
The framework represents a turning point for AI oversight. While Trump's administration has adopted a light-touch regulatory approach, rapid advances in AI models for cybersecurity and hacking are beginning to upend it.
It's an open question how the new process will shape up and affect the release and adoption of new AI models. The New York Times reported on Tuesday that White House officials said at the meeting that the federal government plans to review only certain types of AI models from companies like Anthropic and OpenAI for potential security risks and does not plan to review open-source AI models so far, though these plans could change as the technology advances. Open-source models allow anyone to download or modify the code for free.
An executive order in early June directed officials to develop a "benchmarking process to assess the advanced cyber capabilities of AI models" within 60 days. The deadline has passed.
The administration's AI oversight approach has been defined by its light-touch regulatory stance. Its AI Action Plan, released in 2025, said that regulation might end up "paralyzing" the industry, and called on agencies to "dismantle unnecessary regulatory barriers that hinder the private sector."
Take a smarter break in your day – and see how far you get.
The executive order in June acknowledged a new reality in which AI models from Anthropic and OpenAI have shown marked improvements in coding, finding cybersecurity exploits, and working on longer-term tasks.
The framework, as sketched in the June order, will require companies to share their AI models with the government for up to 30 days before releasing them to the public, giving the government a head start in assessing their cyber capabilities.
Crucially, Trump's executive order described the framework as "voluntary," though it's an open question how the administration would treat companies that skip the process. Jessica Ji, an analyst at Georgetown's Center for Security and Emerging Technology, said it seems like officials are trying to create a "soft compliance regime."
"Ideally, they can say we're not regulating, this is a voluntary program, but all of the companies whose models they think present potential national security risks would comply," she said.
Regulating AI is tricky. Rules could target the top-tier labs, specific products, or individual AI models. If regulators home in on the models, more questions pop up. Do they base their rules on the models' size, or when they're released, or what they're capable of?
The conversation in D.C. over recent months has turned around the AI models' capabilities because of new developments like Anthropic's Mythos model and OpenAI's security incident with the AI company Hugging Face.
Trump's executive order asked his agencies to develop a definition of the "covered frontier models" that would be vetted by his administration. Chinese open-source and open-weight models, which have sparked furor in the industry recently because of their advanced capabilities, are almost certainly part of the Tuesday conversation, Ji said.
She added that figuring out ways to categorize AI models is a crucial step for the rulemakers.
It's unclear so far how the framework will affect the rollout of AI models, especially if it truly feels voluntary to companies. Ji said she's curious whether there will be any clear guidance on what the administration can do if it finds a dangerous capability while testing unreleased technology.
Source: Read the original article on www.businessinsider.com