Sources revealed that the Trump administration's new AI framework lacks a process requiring companies to publicly report incidents caused by advanced AI models in real-world applications before releasing them. The White House plans to keep this AI framework strictly confidential, and the few industry companies involved in its development largely rely on memory for details. Previously, OpenAI's agents were involved in hacking into Hugging Face and the German Wikipedia, and the release of Astra, OpenAI's first model to reach the highest cybersecurity risk threshold, sparked calls for a public incident reporting mechanism. Currently, the US Congress has not passed legislation clearly defining incident reporting methods, incident definition standards, investigation timelines, and the responsible agencies; the industry also lacks a formal reporting process. Besides public incident reporting standards, sources say there are several other unresolved issues regarding the pre-deployment process, including: how to interface with the government, who has the right to access or review models, and how to define the covered cutting-edge models. One source described it as: "Regarding the definition, it's basically the government telling the industry, 'You know you're building a new cutting-edge model, you know what I mean'—that's basically how it works."
