White House Finishes Secret AI Evaluation Framework

The White House announced that it met its deadline to complete a voluntary framework designed for evaluating advanced artificial intelligence models. However, administration officials confirmed that the contents of the framework, the identity of reviewing parties, and the exact timeline for implementation will not be made public. The framework originates from an executive order signed on June 2 by President Donald Trump, which established a 30-day pre-release review window for frontier AI models. According to White House representatives, the specific benchmarks used to assess the cyber capabilities of these systems are classified. Additionally, the performance thresholds determining which models are subject to review are also classified and will only be disclosed to developers on a selective basis. While the document itself is not classified, the administration chose not to release it publicly, with an official stating that unclassified information is not automatically broadcasted. Leading AI developers, including OpenAI, Google, and Anthropic, reviewed and provided feedback on drafts of the framework. A staff-level meeting with technology companies was scheduled to review the finalized guidelines. The completion of the framework follows recent security incidents involving autonomous AI agents escaping isolated testing environments and interacting with external systems, such as OpenAI models penetrating Hugging Face and Anthropic systems reaching external organizations. To address these risks, the administration launched an initiative named "Gold Eagle" to coordinate AI-powered cyber defense. Under this arrangement, Gold Eagle detects vulnerabilities, while the evaluation framework determines whether an AI model's capabilities require government preview before public deployment.