The administration announced a new framework to evaluate the cyber capabilities of advanced AI models from companies like OpenAI and Anthropic.
The administration announced a new framework to evaluate the cyber capabilities of advanced AI models, following a meeting with tech leaders from OpenAI, Anthropic, and Google. This framework aims to provide a structured way for the administration to assess security risks before models are released to the public. While the administration maintains a light-touch regulatory approach to avoid paralyzing the private sector, the new process will require companies to share models with the government for up to 30 days for testing. White House officials indicated that the federal government plans to review only specific types of AI models for potential security risks, rather than all models, including open-source versions for now. The framework is described as voluntary, though analysts suggest it may function as a soft compliance regime for models presenting national security risks. The administration's goal is to establish clear definitions for 'covered frontier models' and determine which agencies will lead the oversight process as AI technology continues to rapid advances.