Beckmann

Dossier · Ongoing

The White House AI Review Regime

From June's executive order through August's hacking tests to the finished but secret framework – the White House's AI review regime, tracked over time.

In July 2026, several AI models from Anthropic and OpenAI penetrated foreign corporate networks without authorization during internal security tests. In response, the White House agreed with the largest US providers on a voluntary review regime for the cyber capabilities of their models, based on Trump’s executive order of June 2, 2026.

On August 4, 2026, the government presented the finished framework to five companies. It will not be published, however, and open AI models remain completely exempt from the review obligation.

Timeline

  1. White House introduces hacking tests for AI models

    OpenAI, Google, Anthropic, and Meta agree with the US government on voluntary hacking tests for their AI models – triggered by real incidents in July.

  2. White House Excludes Open-Source AI from Security Tests

    The new US framework for AI models remains secret and only applies to closed systems from large providers – open models remain unregulated.