White House Intros Classified Cybersecurity Review for Frontier AI Models
The White House has announced it is moving forward with a voluntary framework for testing the cybersecurity capabilities and risks of advanced artificial intelligence models, though the standards guiding those evaluations remain classified.
Representatives from OpenAI, Anthropic, Google, Meta, Nvidia and other leading AI companies recently met with administration officials to discuss the framework. The White House did not announce any agreements following the meeting or publicly disclose how models will be evaluated before their broader release.
The framework would allow developers to provide the federal government with early access to models that demonstrate advanced cybersecurity capabilities. Government agencies would then evaluate the models using a classified benchmarking process before they are made more widely available.
The initiative marks one of the Trump administration’s most substantial efforts to establish a formal review process for frontier AI models. It also exposes a central tension in U.S. AI policy: how to evaluate systems that may strengthen cyber defenses while also making sophisticated offensive capabilities more accessible.
A Framework Built Behind Closed Doors
President Donald Trump directed federal agencies to create the framework in a June executive order.
Under the order, developers can work with the government to determine whether a model qualifies as a “covered frontier model.” Participating companies can provide government evaluators with access for up to 30 days before the model is released to other trusted partners.
The order requires confidentiality, cybersecurity, intellectual property, insider risk, and nondisclosure protections for models submitted for review. It also explicitly states that the program does not establish a mandatory licensing, preclearance, or permitting system for AI model releases.
The White House said that it had completed the framework by its deadline.
“Discussions with industry about next steps are underway,” a White House official told Axios.
The administration has not released the framework, identified the models likely to fall under it, or explained when testing will begin. The benchmark used to determine whether a model has sufficiently advanced cyber capabilities is classified.
That secrecy may be partly unavoidable. Publishing detailed offensive cybersecurity benchmarks could give malicious actors a roadmap for evaluating or improving their own systems. But keeping the entire process confidential also makes it difficult for customers, smaller developers, researchers, and policymakers to assess whether the rules are being applied consistently.
What Emerged After the Meeting
No participant has publicly described a final agreement from the White House meeting. However, according to reporting by Politico, OpenAI, Anthropic, and Google had previously reviewed a draft version of the framework and submitted joint feedback to the White House. The companies reportedly advocated that developers should be allowed to continue conducting A/B testing during model development without government interference, a position the White House accepted.
Although the administration has not publicly confirmed specific conditions regarding federal funding or procurement leverage, the Office of Science and Technology Policy is actively working on testing standards. The White House is also finalizing the roles that agencies such as the National Institute of Standards and Technology and the Cybersecurity and Infrastructure Security Agency will play in evaluating these models.
