News

White House puts a voluntary frontier-model testing framework on the table

In the first days of August 2026 the White House gathered the largest American AI developers — OpenAI, Anthropic, Google and Meta — to walk them through a finished framework for government safety testing of frontier models. The document itself has not been published. What is known about the mechanism comes from reporting rather than from an official text. Participation is opt-in: a developer may hand a model over to the government for evaluation for up to 30 days before public release, and the framework may also stipulate which trusted partners get early access. It is explicitly not a licensing regime — the administration has ruled out turning it into mandatory preclearance, and there is no enforcement mechanism behind it. The framework grew out of a June 2026 executive order on AI and cybersecurity, which told the administration's cyber team to build tests of whether American frontier models can facilitate cyberattacks. Who will actually run the evaluations is still unsettled: some reporting points to the Center for AI Standards and Innovation (CAISI) at the Commerce Department, other accounts to the Office of the National Cyber Director, with the NSA also named as a candidate. For the catalogue this matters in a practical way. If the largest laboratories join in, a 30-day pre-release evaluation window becomes part of the release calendar for models such as the Claude, GPT and Gemini families — a step between "model finished" and "model available" that until now did not exist in the United States.