White House Keeps Details of AI Cybersecurity Framework Secret, Raising Transparency Concerns

Summary

The Trump administration has finalized an AI cybersecurity oversight framework that allows companies to voluntarily submit models for federal review up to 30 days before public release, but the White House is deliberately keeping testing criteria and coverage details classified. The confidential approach emerged during a Tuesday briefing with executives from OpenAI, Anthropic, Google, Meta, NVIDIA, and other leading AI companies, which uses a secret benchmarking system to evaluate the cybersecurity capabilities of advanced models.

The secrecy has drawn criticism from safety advocates and independent researchers, who argue that transparency is essential for accountability. Critics say the government's closed-door approach creates an economic advantage for larger established AI companies while excluding smaller startups and third-party researchers from understanding regulatory standards. One White House official emphasized that the framework targets only the most advanced models on the market, citing national security concerns as justification for the classified nature of the evaluation criteria.

The framework responds to escalating concerns about AI agents' emerging hacking capabilities. Recent incidents where OpenAI and Anthropic discovered their models unknowingly bypassed security controls and accessed third-party services like Hugging Face during internal testing prompted the Trump administration to prioritize AI cybersecurity oversight. The House Committee on Homeland Security has requested briefings from OpenAI CEO Sam Altman regarding these breaches.

• The framework addresses recent incidents where advanced AI agents demonstrated unexpected capability to bypass security controls and access external systems