As AI cybersecurity fears rise—and lawmakers press tech companies to address them—White House officials welcomed representatives of top AI companies on Tuesday to discuss a new voluntary prerelease review process. The framework will determine how the Trump administration reviews security risks in America’s leading “frontier” AI models before they are made publicly available, and marks a shift for an administration that started as an AI booster, favoring light regulation to help U.S. companies better compete with Chinese AI firms.
But that framework isn’t being published: Only the technology companies taking part in the process, including Anthropic, OpenAI, Microsoft, Meta, Google, and Nvidia, will be privy to the details. And from what has leaked to the press, the framework doesn’t seem to address many of the core concerns about AI safety.
“To the extent that there might be potentially harmful capabilities that [AI] companies are building or testing, and then are being used by governments or entities with access to these advanced models, we wouldn’t know,” Aya Ibrahim, a senior fellow for economic and national security at the AI Now Institute, told TMD. The framework isn’t public, and neither are the government’s evaluations, so the public “can’t scrutinize the quality or the relevance of the benchmarks that these companies are meeting or not,” Ibrahim said. And there’s “no visibility into what firms, what organizations, or actors are participating in shaping this framework.”
Read the full article here.
Research Areas