Histoire
août 5, 2026
White House Keeps AI Rules Secret While Letting Open Models Slip Free
Trump’s administration has completed a voluntary AI pre-release testing framework, but its decision to keep the rules private and exempt open models is leaving outside companies, researchers and allies without answers.
The White House has finished its new playbook for testing powerful AI before release, yet the people expected to live with its consequences may never see the rules. The framework’s narrow focus on closed models has sharpened an already fierce fight over how Washington should police fast-moving AI.
The process began with President Trump’s June executive order, which asked developers to share frontier systems with the government before release to assess cyber risks. The order made the capability-benchmarking process classified, but did not require the broader voluntary framework to be kept secret.1
By Monday, the administration said it had met its deadline. “The voluntary framework outlined in the June 2nd executive order was complete by the deadline,” a White House official said, adding that industry discussions on next steps were under way.2 The official defended the limited disclosure bluntly: “Just because things are unclassified that doesn't mean we are going to broadcast them to everyone.”2
On Tuesday, officials reviewed the framework with industry participants, including leading frontier labs. But companies not invited, as well as policymakers, researchers and U.S. allies, were left guessing at its scope and the identity of any “trusted partners” granted early model access.1 That secrecy matters because the framework is expected to govern up to 30 days of government access before a model’s release, with rules on cybersecurity, insider risks, intellectual property and nondisclosure.2
The most consequential line may be what the framework leaves out. Sources say it covers only closed-source, state-of-the-art models carrying national-security risks—terms it does not clearly define—and explicitly excludes open models. Nothing in the policy should be read as restricting an open model after release.3 Supporters of open development will see a reprieve; critics will see a regulatory blind spot, particularly as cyber threats and Chinese AI advances intensify. Yann LeCun amplified the news in a repost describing the open-model exemption as a “SITUATION UPDATE.”
4
For now, participation remains voluntary. Yet companies seeking to avoid a collision with Washington have strong incentives to comply—while smaller developers, excluded from the room and unclear on the thresholds, are left to decipher a policy they cannot read.