Trump administration urges OpenAI to slow GPT 5.6 release

The Trump administration asked OpenAI to slow the release of GPT 5.6 over security concerns, and Sam Altman told staff the model would go out first in a limited preview to a small group of partners. Customer access would be approved one by one during that preview period, instead of opening the system to a full public launch.

The request came through the Office of the National Cyber Director and the Office of Science and Technology Policy, underscoring how national security has become part of the release calculus for frontier AI. The Office of the National Cyber Director says it coordinates national cyber policy, including information security, threat deterrence and the security of information and communications technology, while a June 5 presidential memorandum on AI in the national security enterprise explicitly included the office in the interagency process.

The pressure lands inside a broader White House framework that favors both acceleration and restraint. On June 2, President Donald J. Trump signed an executive order titled “Promoting Advanced Artificial Intelligence Innovation and Security,” which said the United States would work with the private sector to promote AI innovation and security while hardening systems against external threats. That split objective helps explain why a model can be treated not just as a commercial product, but as a national-security issue before it reaches broad public release.

OpenAI has already been building more controlled distribution channels. In June 2026, the company launched the OpenAI Partner Network and said it would invest $150 million to support enterprise AI adoption. It has also used limited-release access for cybersecurity-focused work, including GPT-5.5-Cyber and the Daybreak cyber partner program, both designed around trusted access to defensive tools.

That approach mirrors a wider pattern in the frontier-AI market, where companies are increasingly segmenting access before a full launch. OpenAI publishes deployment safety materials and system cards through its Deployment Safety Hub, giving the company a formal structure for staged rollouts and risk review. Anthropic has used a similar controlled-release approach for a cybersecurity-capable model, showing that gated access is becoming a standard response when models are powerful enough to raise security questions before they are widely deployed.

The issue now is less about one delayed product than about who gets to slow frontier AI in the United States. If the government can press for a narrower release on security grounds, and a company can answer by moving to a partner-only preview, the market is moving toward a de facto approval process in which advanced models face policy scrutiny before broad distribution.