On a Tuesday in late summer, representatives from six of the most powerful technology companies on Earth — OpenAI, Anthropic, Meta, Google, Nvidia and Microsoft — filed into a private meeting at the White House. The subject was how America plans to check whether new artificial intelligence models pose a threat to global cybersecurity. What was decided inside that room, however, is not something the rest of the world is meant to know. The Trump administration has finalised a voluntary framework requiring AI companies to submit new models for government review up to 30 days before public release.

Yet the White House has confirmed it will not publish the criteria used to assess those models, sharing the details only with a select circle of favoured firms. That opacity is significant. Without published benchmarks, outside researchers cannot verify whether the tests are rigorous, foreign governments cannot assess whether risks to shared digital infrastructure are being managed, and smaller companies that are excluded from the private briefings face an uneven playing field — with no map to the rules they are expected to follow. The framework's origins lie in a quiet geopolitical alarm raised earlier this year.

When Anthropic withheld its Mythos model from public release in April, citing fears it could be used to attack financial and IT systems, the episode forced even a famously hands-off White House to reconsider its approach to AI oversight. An executive order in June established the voluntary review process, though it fell short of mandatory vetting after prominent tech figures reportedly lobbied the president directly against compulsion.