White House asks OpenAI to delay new model over safety concerns
▼ Summary
– OpenAI’s GPT 5.6 will have a limited release to select partners instead of the public, following government requests from the Trump administration.
– CEO Sam Altman told staff that the government will approve access customer by customer during a preview, with a broader release possible weeks later.
– The Trump administration, initially taking a “hands off” approach, now pushes for federal oversight, including an executive order for voluntary testing of new AI models.
– Anthropic previously limited its Claude Mythos model to partners through Project Glasswing, citing risks of misuse, which sparked debate over whether this is genuine safety or marketing.
– Frontier cyber models like Mythos can identify and exploit software vulnerabilities faster than humans, posing risks to organizations, but their closed nature makes threat assessment difficult.
OpenAI’s latest model, GPT 5.6, will not see a typical public launch. Instead, the company plans to release it only to a small circle of trusted partners, following directives from the Trump administration, according to a report from The Information.
During a meeting this week, CEO Sam Altman reportedly informed staff that the government would be “approving access customer by customer” throughout a preview phase. He added that if the limited rollout proceeds smoothly, OpenAI hopes to expand access to the general public “a couple of weeks later.”
This marks a notable shift for the Trump administration, which initially favored a “hands off” approach to AI regulation. In recent months, however, the White House has pushed for greater federal oversight of emerging models. Earlier this month, Trump signed an executive order urging certain AI companies to voluntarily submit new models for government testing and evaluation before any public release.
The agencies reportedly behind the request for a restricted release include the Office of the National Cyber Director and the Office of Science and Technology Policy. According to The Information, OpenAI staffers “worked closely” with these offices to shape the upcoming deployment.
In essence, the administration appears to be nudging OpenAI toward the same cautious stance that Anthropic has already adopted voluntarily. Earlier this year, Anthropic sparked debate by releasing its new frontier cyber model, Claude Mythos, exclusively through a program called Project Glasswing, accessible only to a select group of partners. The company argued the model was too powerful and could, if misused, cause significant harm. Critics have since questioned whether this was a genuine safety measure or a clever marketing strategy. The truth likely lies somewhere in between.
The broader concern is that frontier cyber tools like Mythos can identify and exploit software vulnerabilities at speeds no human analyst can match. Many enterprise systems contain hidden bugs that serve as entry points for attackers, making such tools a serious threat to organizations running complex software infrastructure. Yet because these models remain closed to the public, it is difficult to assess their true danger.
Cybercriminals have long used automated tools, but the rise of generative AI has amplified their capabilities. Large language models are now adept at writing malware, and some can even execute entire ransomware attacks autonomously. The Trump administration’s push for limited releases reflects growing unease about the potential for powerful AI to fall into the wrong hands.
(Source: TechCrunch)
