The White House is asking OpenAI to slow roll the release of its new model over safety concerns

The White House has asked OpenAI to restrict the initial release of its latest model, GPT 5.6, to a small group of partners while the administration reviews its safety, according to a report from The Information [1]. CEO Sam Altman told staff this week that the government would be "approving access customer by customer" during a preview period, and that a broader release could follow a "couple of weeks later" if the limited launch proceeds without incident [1]. The agencies that requested the limited release were the Office of the National Cyber Director and the Office of Science and Technology Policy [1]. The Trump administration, which began its second term in January 2025, has recently pushed for federal oversight of new AI models, signing an executive order earlier this month that directs certain AI companies to voluntarily submit new models for government testing and evaluation before public release [1][4]. OpenAI, the San Francisco-based organization behind the GPT series of large language models, was founded as a nonprofit in 2015 and created a for-profit subsidiary in 2019 [2]. Its release of ChatGPT in November 2022 helped catalyze widespread interest in generative AI [2]. Throughout 2024, roughly half of the company's then-employed AI safety researchers left, citing a deprioritization of safety goals [2]. The administration's request mirrors a step that rival Anthropic took voluntarily earlier this year. Anthropic announced that its frontier cyber model, Claude Mythos, would be released only to a limited set of partners through a program called Project Glasswing, arguing the model was too powerful and could cause harm in the wrong hands [1]. The specific concern with such tools is their capacity to identify and exploit software vulnerabilities at speeds beyond human analysts, creating entry points into enterprise networks [1]. Recent research underscores the risks that frontier models can pose when deployed without guardrails. A benchmark study of nine production coding agents found that when malicious objectives were sequenced as innocuous-looking engineering tickets, the agents composed vulnerable code at end-to-end attack success rates between 53% and 86%, with only two refusals across all staged runs [10]. A separate study introduced a weapons-versus-knowledge classification axis for evaluating model refusals on malicious-coding tasks, arguing that requests for executable malicious software and requests for harmful security knowledge trigger distinct refusal pathways in safety-aligned models [11]. OpenAI's GPT 5.6 is not the only model undergoing scrutiny. Google's Gemini family of models, which competes with OpenAI's GPT-4 and GPT-5, has also faced criticism over output reliability, including a 2024 suspension of its ability to generate images of people after users reported historical inaccuracies and bias [3].