White House Prepares Joint Rules for Early Access to Most Powerful AI Models

The Trump administration is wrapping up negotiations with OpenAI, Google, and Anthropic over joint, voluntary rules governing the release of the most advanced artificial intelligence models. The framework would give the U.S. government early visibility into new models before their public launch, and an announcement could come within days.

What's Set to Change

Executive Order 14409, titled 'Promoting Advanced Artificial Intelligence Innovation and Security,' requires government agencies to develop, within 60 days, a voluntary mechanism through which developers of the most powerful models will be able, or encouraged, to share them with the government ahead of release. In practice, this means models classified as 'covered frontier models' will first undergo a security review before reaching trusted business partners and foreign customers.

A key element of the framework is a classified benchmarking process designed to assess models' advanced cyber capabilities. The decision on whether a given model qualifies as a frontier model requiring advance notification would rest with the director of the National Security Agency, in consultation with other bodies responsible for cybersecurity and defense. The framework also envisions tiered access, with different terms for domestic partners than for commercial customers abroad.

Recent Precedents

Before the framework was formally announced, the administration had already applied it in practice twice. OpenAI had to delay the full public release of GPT-5.6, also known as Sol, at the government's explicit request. During the limited testing period, only about 20 customers were granted access, each personally approved by Commerce Secretary Howard Lutnick. OpenAI publicly stressed that its compliance with the restriction was voluntary rather than mandated.

The second case involved a dispute over Anthropic, whose models were subject to binding Commerce Department export restrictions over national security concerns tied to a South Korean telecom operator with ties to China. The restrictions were lifted on July 1, but only after Anthropic agreed to implement filters blocking attempts to break the model's safeguards more than 99 percent of the time. It's the first known instance of the U.S. government imposing a specific technical condition as the price for approving the export of an AI model.

Criticism and Concerns

The executive order explicitly states that it does not establish a mandatory licensing system, pre-certification, or approval requirement for developing new AI models. Critics counter that in practice, the government is using existing legal authority to build a mechanism that functions like licensing, even if it isn't formally one.

We're seeing the government turn existing legal authority into what is, in practice, a hidden licensing system - analysis by the Accountable AI Lab, Wharton School, cited by Fortune

Tech companies have spent months lobbying to keep new regulations from slowing the pace of AI deployment in the U.S., driven largely by fears of competition from China. At the same time, the Trump administration had previously assured that it had no intention of introducing rules that would curb the sector's innovation. The current framework is thus an attempt to reconcile two competing goals: keeping up the pace of model development while retaining control over who gets access to the most powerful tools, and when.

Implications for Business and the Market

For businesses relying on OpenAI, Google, and Anthropic models, this means the launch of cutting-edge models could now be delayed by additional weeks before they reach broad commercial use, including outside the United States. Companies planning deployments built on the latest model versions should expect that access may be contingent on approval from U.S. authorities, particularly for models with heightened cybersecurity-relevant capabilities.

The final rules are expected to be announced within the coming week, though neither side has officially confirmed an exact date. The formal deadline set in the executive order is August 1, 2026, the same day the European Union's requirement to label AI-generated content takes effect, meaning the global AI industry will enter August facing two new oversight regimes running in parallel on opposite sides of the Atlantic.

Sources: Fortune (fortune.com), AI Weekly (aiweekly.co), The White House (whitehouse.gov)