OpenAI is reportedly taking a different approach with the launch of its upcoming GPT-5.6 model after the U.S. government raised safety concerns. Rather than making the model immediately available to the public, the company is expected to provide access only to a limited group of trusted partners during an initial preview period, according to a report by The Information.
During a company meeting this week, CEO Sam Altman reportedly told employees that customer access would be approved individually by the government throughout the preview phase. If the rollout proceeds without major issues, OpenAI hopes to expand availability to a wider audience a few weeks later.
The strategy suggests the Trump administration is encouraging OpenAI to adopt a more cautious release process, similar to Anthropic’s approach of restricting access to its most advanced AI models.
According to The Information, OpenAI has worked closely with U.S. officials ahead of the release. The report says both the Office of the National Cyber Director and the Office of Science and Technology Policy requested that the initial rollout remain limited while the model undergoes additional evaluation.
Although the Trump administration initially promoted a relatively hands-off approach to artificial intelligence, its stance has evolved in recent months. Earlier this month, President Donald Trump signed an executive order encouraging certain AI developers to voluntarily submit advanced models for government testing and evaluation before broader public deployment.
Anthropic faced similar attention earlier this year after limiting access to its frontier cybersecurity model, Claude Mythos, through its Project Glasswing program. The company argued that the model’s capabilities could pose significant risks if widely released before appropriate safeguards were in place. Critics questioned whether the move reflected genuine safety concerns or a marketing strategy, while others viewed it as a combination of both.
The debate comes as generative AI continues to enhance cyber capabilities. While cybercriminals have long relied on automation, advanced large language models can now assist in writing malware and, in some cases, help automate portions of sophisticated ransomware attacks.
Security experts are particularly concerned about frontier AI systems capable of identifying and exploiting software vulnerabilities far more quickly than human researchers. Because modern software often contains hidden flaws that can serve as entry points for attackers, these capabilities could create serious cybersecurity risks for governments, businesses, and critical infrastructure.
At the same time, the actual capabilities of these frontier models remain difficult to independently verify, as access continues to be restricted to a small number of approved organizations.
