White House finalizes voluntary cybersecurity test framework for frontier AI models

The White House has reportedly finished a voluntary framework for testing whether frontier artificial-intelligence models can be used to find software vulnerabilities or carry out cyberattacks. It convened major AI companies on August 3, 2026 to review the document.

The framework has not been published, and it remains unclear whether it will be, an unusual level of secrecy for a process that could shape how the most capable AI systems reach the public.

The Office of the National Cyber Director, the White House office responsible for cyber policy, is said to be hosting the meeting, with staff-level representatives from OpenAI, Anthropic, Google, Microsoft and xAI expected to attend. The framework reportedly grows out of President Trump’s June 2, 2026 executive order and would let companies give the federal government up to 30 days of early access to certain frontier models before a wider release, so officials can assess whether a model could discover exploitable flaws or automate an attack.

The administration has said the arrangement cannot be turned into a mandatory licensing or preclearance regime, keeping participation voluntary. That framing matters: it leaves labs free to opt out, and at least one has. Meta has reportedly declined to join the underlying pre-release evaluation process, known as TRAINS, for Testing Risks of AI for National Security, even as it is listed among the companies expected at the meeting itself.

Voluntary early-access schemes have a mixed record. Without published criteria, it is hard to judge what testing means in practice, who sees the results, or what happens if a model fails a review. The value of a 30-day window also depends on whether labs hand over the versions they actually ship rather than an earlier checkpoint.

The framework’s real reach will become clear only if the administration publishes it or the participating labs describe what they agreed to.