White House Orders OpenAI to Limit GPT-5.6 Release to Approved Partners Only

The Trump administration has directed OpenAI to restrict the release of its newest model, GPT-5.6, to a select group of government-approved partners rather than launching it publicly, according to The Information.

CEO Sam Altman told staff at a meeting this week that the government would be "approving access customer by customer" during a preview period. If the limited release proceeds without issues, OpenAI hopes to follow with a broader public release "a couple of weeks later," according to Altman.

Government Agencies Behind the Decision

The Office of the National Cyber Director and the Office of Science and Technology Policy requested the limited release approach. OpenAI staffers reportedly "worked closely" with the government on the upcoming release and review process.

This marks a shift from the Trump administration's initial "hands off" positioning on AI regulation. Earlier in June, Trump signed an executive order directing certain AI companies to voluntarily submit new models to the government for testing and evaluation before public release.

Following Anthropic's Playbook

The restricted release mirrors Anthropic's approach with Claude Mythos, its frontier cyber model released earlier this year only to select partners through Project Glasswing. Anthropic stated that Mythos was too powerful to release publicly due to its capabilities in identifying and exploiting software vulnerabilities at speeds exceeding human analysts.

The concern centers on frontier cyber models' ability to write malware and execute ransomware attacks autonomously. These models can identify hidden bugs in software systems that serve as entry points into enterprise networks—a significant threat to organizations running complex infrastructure.

Debate Over Legitimate Threat vs. Marketing

Anthropic's restricted release of Claude Mythos sparked debate over whether such safety concerns represent genuine risk mitigation or marketing strategy. With both models remaining closed to the public, independent verification of their threat level remains difficult.

LLMs have demonstrated capability in writing malware, and cybercriminals have adopted automated tools powered by generative AI. However, the actual capabilities of restricted frontier models like Mythos and GPT-5.6 cannot be independently assessed without broader access.

What This Means

This marks the first time the U.S. government has explicitly directed a major AI lab to restrict model access, moving beyond voluntary frameworks to direct intervention. The customer-by-customer approval process establishes a precedent for government gatekeeping of AI model releases, particularly for models with potential cybersecurity applications. Whether this approach effectively balances security concerns against innovation and transparency remains an open question, especially given the lack of public evidence about these models' actual capabilities.