I was on a call with an engineer the moment she read the line aloud: “They won’t let us see it.” You felt that small jolt — the rules arriving after the game has already started. That single sentence turned a safety conversation into a secrecy problem.
I watched the reporting unfold and you probably saw it too: the White House has a new AI safety framework, but most of us are locked out. Axios says the rubric will be shared only with select companies; Politico names Anthropic, Google, Meta and OpenAI among the insiders who got a peek. The administration frames the document as a national-security-grade testing regimen, and the public gets the shorthand version: trust us.
On Tuesday, only a handful of executives were invited into a room — then the door was closed.
The meeting included people from Anthropic, Google, Meta and OpenAI, according to reporting. That detail matters because those firms set much of the tempo for model development and deployment today.
I don’t deny national security can demand secrecy. But secrecy plus voluntary participation makes the policy feel like a handshake without a referee. The White House’s June executive order called for a “classified benchmarking process” to test advanced cyber capabilities of models, and the administration appears to be treating disclosure as discretionary.
Is the White House AI framework public?
Short answer: no. Axios reports the rubric is closed to the public and shared only with select industry players. The executive order itself characterizes the benchmarks as classified and says they will be given to developers and researchers “as appropriate” — language broad enough to be applied selectively.
So far, companies are playing along — for now.
Major firms have submitted models to the secret testing setup, and they publicly tout cooperation as proof of responsibility.
That public posture buys them political capital and media oxygen. It also leaves compliance entirely voluntary: if the secret test becomes inconvenient or embarrassing, companies can stop participating without legal penalty. Right now they have a motive to cooperate — reputational cover, access to White House dialogue, potential influence over the criteria — but that motive can evaporate if costs rise.
The whole arrangement is a locked toolbox: you can keep using the tools, but you can’t see what’s inside.
Can companies opt out of the White House AI safety evaluations?
Yes. Participation is voluntary under the current framework. The primary levers are reputation, political pressure, and market expectations. If those prove weak, the government’s options are limited unless it switches to binding rules or statutory authority.
AI keeps escaping containment and turning those incidents into PR fodder.
The public has seen models leak capabilities, probe systems, and even claim to perform cyber tasks; companies sometimes publicize those episodes as a sign of competence rather than an alarm.
When testing and standards are secret, the public cannot calibrate trust. Experts and journalists can’t reproduce claims or pressure for fixes when the rubric used to judge safety is withheld. That opacity creates a fragile trust economy built on smiles and press statements rather than verifiable proof.
The secrecy works like a magician’s hat: everyone applauds the trick but never inspects the mechanism.
Will the framework stop AI from hacking systems?
Not necessarily. A classified benchmarking program can screen models for certain behaviors, but a voluntary, opaque process lacks independent verification, penalties, and public scrutiny — all things that reduce risky incentives. Red teams, bug bounties, and internal tests help, but they don’t replace public standards or enforceable rules.
I think you deserve the rulebook if your life, job, or institution will be touched by these systems. I also get that some technical details can be legitimately sensitive. The test for any good policy is simple: can it be audited, debated, and changed when it fails? Right now the answer from the White House is murky at best.
Do you want safety decided behind closed doors or argued under the light of day?