Back in June, the White House issued an executive order telling federal agencies and the top AI labs to work out a framework for testing the most powerful new models before they go public. The deadline was August 1st. They hit it. Then they did something odd. They decided not to tell anyone what is actually in it.
Officials from Google, OpenAI, Anthropic, Meta and a handful of smaller labs sat down with the government to hash out voluntary testing rules. Part of that framework reportedly includes classified benchmarks that decide which models count as covered frontier models in the first place. A company could cross that line without ever knowing where the line was drawn.
Smaller AI labs are the ones feeling this the hardest. The giants have compliance teams and government relations staff who can read between the lines and guess what regulators want. A ten person startup building a frontier model does not have that luxury. If the rulebook stays sealed, only the companies big enough to already have a seat at the table get to prepare for it.
I get the instinct here. Some of what sits inside a framework like this probably touches on real security concerns, the kind of thing you do not want printed in a press release for anyone to game. But there is a difference between keeping specific attack details quiet and keeping the entire rulebook hidden from the companies it applies to.
Good regulation usually works because everyone can see the rules and plan around them. Secret rules mean the first time you find out you crossed a line is when someone taps you on the shoulder about it.
None of this means frontier AI oversight is a bad idea. It is a good one, maybe overdue. I just want the people building these models, big and small, to actually know what they are being held to. Transparency was supposed to be the whole point of a safety framework in the first place.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://fortune.com/2026/08/04/baffling-white-house-wont-publicly-release-ai-model-evaluation-framework-it-reviewed-today-with-openai-anthropic-microsoft-and-others/
https://fortune.com/2026/08/06/washington-is-keeping-its-ai-rulebook-private-smaller-ai-labs-arent-happy/
https://www.defenseone.com/technology/2026/08/ai-models-white-house-and-companies-secret-safety-measures/415227/
https://siliconangle.com/2026/08/04/white-house-ai-firms-keep-safety-framework-talks-private/
Repost this. Thanks.