The White House keeps its AI safety test under wraps

Just days after OpenAI and Anthropic disclosed that their AI agents had breached systems at other companies, Washington called the labs in for a sit-down. But whatever rules they are agreeing on behind closed doors, you are not allowed to see them.

Three sources told Axios that the White House will not publicly release its new framework for evaluating advanced AI models. Outlined in a June 2 executive order from the Trump administration, the framework governs how AI developers can voluntarily work with the government to assess frontier models for cyber capabilities up to 30 days before public release.

Classified tests and excluded allies

Only a handful of companies invited to staff-level meetings know what is actually inside the document. According to reports from Reuters and Axios, attendees included Meta, Google, Anthropic, OpenAI, and Nvidia. Independent researchers, civil society, policymakers, and U.S. allies were completely shut out, and the cyber-benchmarking process itself has been classified.

On Tuesday, Trump advisers also told developers that the voluntary tests will not cover open-weight models like Meta’s Llama or Nvidia’s Nemotron. The administration is focusing exclusively on closed frontier systems from labs like OpenAI, Google, and Anthropic.

The push comes amid mounting political pressure. Five Democratic senators sent a letter calling for permanent federal legislation for frontier safety, warning against an opaque, case-by-case regulatory regime. Meanwhile, 15 Republican state attorneys general launched an inquiry into an OpenAI agent breach involving Hugging Face, demanding the lab preserve all records.

Why it matters

On paper, a 30-day pre-release review could catch dangerous autonomous hacking capabilities before another agent goes rogue in the wild. But we think the execution creates a massive accountability vacuum.

Because the entire framework is voluntary and the testing benchmarks are classified, the public has no way of knowing which labs actually hand over their models or what standard constitutes a pass or a fail. For developers building on these APIs, it offers zero transparent safety assurances.

It also sets up a stark regulatory split. While the European Union enforces binding, transparent requirements under the EU AI Act, Washington is leaning into an informal, closed-door gentlemen’s agreement with a few domestic titans. Exempting open-weight models will please the open-source community, but keeping safety evaluations secret leaves enterprises and international allies entirely in the dark.