OpenAI to Limit GPT 5.6 Release Following Government Safety Review

OpenAI plans to restrict access to its latest model, GPT 5.6, sharing it only with a select group of partners rather than releasing it publicly. This decision comes after the Trump administration requested a controlled rollout, according to The Information.

During a meeting this week, CEO Sam Altman informed staff that the government would be “approving access customer by customer” throughout the preview phase. Altman also indicated that if the limited release proceeds smoothly, OpenAI aims to expand availability to a broader audience within a few weeks.

This approach aligns with the strategy already adopted by Anthropic, which has voluntarily kept its most advanced AI models confidential. The Information reports that OpenAI collaborated closely with government agencies, including the Office of the National Cyber Director and the Office of Science and Technology Policy, in preparing for the model’s release.

Although the Trump administration initially advocated a “hands off” stance toward AI, it has recently intensified calls for federal oversight of new models. Earlier this month, Trump signed an executive order encouraging certain AI companies to voluntarily submit new models for government testing and evaluation prior to public release.

Anthropic stirred debate earlier this year by announcing that its frontier cyber model, Claude Mythos, would be accessible only to a limited group of partners through a program named Project Glasswing. The company justified this by stating the model’s power could cause harm if misused. Observers continue to discuss whether this is a marketing tactic or a genuine safety measure.

Cybercriminals have long employed automated tools, but generative AI now provides them with enhanced digital capabilities. Large language models have demonstrated proficiency in writing malware and can even autonomously carry out ransomware attacks.

The primary concern regarding frontier cyber tools like Mythos is their ability to detect and exploit software vulnerabilities at speeds far exceeding human analysts. Since many software systems contain hidden bugs that serve as entry points into enterprise networks, this presents a significant risk for organizations managing complex software infrastructures. However, because these models remain restricted from public access, assessing their true threat level remains challenging.