White House Intervenes: OpenAI Urged to Pause New AI Model Release Over Safety Fears

OpenAI's latest model, GPT 5.6, is reportedly departing from the company's usual public release strategy. Instead, it is slated for a limited distribution to a select group of close partners, a decision reportedly influenced by the Trump administration. At a recent meeting, CEO Sam Altman informed staff that the government would be "approving access customer by customer" during an initial preview period. Should this limited release prove successful, OpenAI hopes to follow with a broader public launch "a couple of weeks later." This move suggests that the Trump administration is exerting pressure on OpenAI to adopt a cautious approach, mirroring the strategy already voluntarily implemented by Anthropic for its most powerful AI models.

According to reports, not only is OpenAI's new model undergoing review by the administration, but company staffers have also collaborated closely with government officials on the upcoming release. The agencies reportedly requesting a limited release include the Office of the National Cyber Director and the Office of Science and Technology Policy. This involvement aligns with a recent shift in the Trump administration's stance on AI; after initially advocating a "hands-off" approach, it has recently pushed for federal oversight of new models. Earlier this month, the administration signed an executive order directing certain AI companies to voluntarily submit new models for government testing and evaluation prior to public release.

This situation draws parallels with Anthropic's earlier decision regarding its frontier cyber model, Claude Mythos. Anthropic generated considerable discussion when it announced that Mythos would only be released to a small group of partners through its "Project Glasswing" program. The company justified this limited release by arguing that the model was exceptionally powerful and carried the potential for significant harm if misused. Observers have since debated whether Anthropic's rationale was a strategic marketing tactic or a genuine effort to prevent the misuse of a potent AI model, with the reality likely lying somewhere in between.

The concerns surrounding powerful AI models, particularly in the realm of cybersecurity, are substantial. Cybercriminals have long utilized automated tools, but the advent of generative AI has provided them with unprecedented digital capabilities. Large Language Models (LLMs) have demonstrated proficiency in writing malware, with some even capable of executing entire ransomware attacks autonomously. The specific apprehension regarding frontier cyber tools like Mythos is their purported ability to identify and exploit software vulnerabilities at speeds far exceeding human capacity. Given that many complex software systems harbor hidden bugs that can serve as entry points into enterprise networks, this capability presents a clear and significant threat to any organization managing intricate software infrastructure.

Ultimately, the true extent of the threat posed by these highly advanced, closed AI models remains challenging to ascertain due to their limited public accessibility. The cautious approach taken by both OpenAI and Anthropic, particularly under government scrutiny, underscores the growing awareness of the potential risks associated with increasingly powerful artificial intelligence.