The artificial intelligence (AI) security checks requested by Washington would reportedly target only OpenAI’s so-called “closed” models, Google, and Anthropic, apparently sparing the “open” models from Meta or the Chinese company DeepSeek, the U.S. media reported on Tuesday, according to AFP.
White House officials presented this confidential framework on Tuesday during a meeting with representatives from OpenAI, Anthropic, Google, Meta, Microsoft, and Nvidia.
The planned oversight would be limited to “closed-source” models “equipped with state-of-the-art capabilities and posing risks to national security,” according to Axios, with nothing in the text intended “to be interpreted as restricting open-source models.”
Closed-source models remain hosted on their creator’s servers, which retain control over their use without disclosing it.
In contrast, users of open-source models can download and modify them. They can remove some of the security measures, turning them into tools for cyberattacks. Nvidia and Meta are developing such models in the United States, Alibaba and Moonshot AI in China, and Mistral in France.
The measure stems from an executive order signed on June 2 by U.S. President Donald Trump, which paved the way for a federal review of the most advanced artificial intelligence models up to 30 days before they are brought to market.
This principle of review—intended to be voluntary—represents a significant shift for the Trump administration, which was previously dominated by opponents of any regulation, accused of stifling innovation in the face of competition from China.
However, this security framework will not be made public, according to reports in the American media. “The rules can only regulate artificial intelligence companies if we, outside of those companies, know what they consist of,” criticized Brad Carson, president of Americans for Responsible Innovation, in a statement.
Companies collaborating with the administration “put innovation, security, and U.S. cybersecurity first,” White House spokesperson Liz Huston told the Wall Street Journal.
In late June, several observers believed that stricter U.S. regulations would make open-source and low-cost models, such as those from DeepSeek, more attractive to foreign customers eager to avoid dependence on Washington.
Anthropic, which was forced by Washington in June to suspend access to two of its models, advocated at the time for “strict regulation,” but one that should be “applied equally to all developers of advanced models.”
This meeting follows a series of incidents in which OpenAI’s most powerful model (GPT-5.6 Sol) and Anthropic’s (Mythos 5) gained unauthorized access to other companies.
On Tuesday, the UK’s Artificial Intelligence Security Institute (AISI) reported that Mythos 5, during an evaluation with safety measures disabled, created a fake identity to contact developers and convince them to embed malware in their software. The institute described this as a deception of unprecedented severity.
AGERPRES
