White House Asks OpenAI to Delay New Model Launch for Security Reasons

Fundamental Shifts in Release Policy

OpenAI's launch of its new model, GPT 5.6, will not follow the company's customary wide-scale release pattern. According to informed sources, the company has decided to restrict access to the model during its initial phase, sharing it with only a limited group of selected partners — a decision made in direct response to a request from the Trump administration.

Direct Instructions from the U.S. Government

At a recent internal meeting, CEO Sam Altman told employees that the government would review and approve access requests on an individual basis during the initial preview period. Altman indicated that if this limited phase proceeds without issues, the company hopes to move toward a broader public launch within a few weeks.

The government agencies that requested the restricted release are:

• The Office of the National Cyber Director
• The Office of Science and Technology Policy

These agencies did not limit their role to review alone — reports indicate that their staff worked closely with the OpenAI team throughout the preparation stages for the upcoming release.

A Shift in the Trump Administration's Stance on AI

This approach marks a notable shift in U.S. administration policy, which has long championed a hands-off approach to the AI sector. However, recent months have seen a significant uptick in federal oversight of new models, culminating in the signing of an executive order requiring certain AI companies to submit their models to the government for testing and evaluation before any public release.

Anthropic Pioneered This Path

What OpenAI is doing today is not entirely new in the tech landscape. Anthropic previously sparked widespread debate when it announced that its advanced cybersecurity model, Claude Mythos, would only be available to a narrow circle of partners through a program it called "Project Glasswing" — citing the model's capabilities as exceeding what could be made available without strict safety controls.

Observers were divided in their assessment: is this a genuine commitment to safety, or simply clever marketing designed to create an aura of exclusivity around the product? The answer most likely falls somewhere in the gray area between the two.

Why Do Cybersecurity Models Raise Such Concern?

Fear of AI tools is nothing new — cybercriminals have been using automated tools for many years. But with the emergence of large language models, the digital arsenal of cyberattacks has become far more dangerous and sophisticated. These models have demonstrated high proficiency in writing malicious code, and some are capable of executing complete ransomware attacks with full autonomy.

The true danger of advanced cybersecurity models like Mythos lies in their alleged ability to:

• Detect and identify software vulnerabilities at speeds that surpass any human analyst
• Exploit these vulnerabilities as entry points into complex enterprise systems
• Execute complete attack chains without human intervention

Given that most software systems contain hidden vulnerabilities that go undetected for extended periods, the existence of AI models capable of uncovering and exploiting them poses a serious threat to any organization relying on a complex digital infrastructure.

The Bottom Line

These developments reveal a growing maturity in how AI risks are handled at the U.S. government level. After a period of largely unrestricted openness, key players — both the government and the companies themselves — are moving toward a more cautious approach to releasing the most powerful and impactful models. The question remains: is this caution sufficient to keep pace with the rapidly accelerating rate of progress in this field?

✦ بقلم فريق دروب أيديا