White House Tells CAISI to Halt Public AI Model Reports

The Center for AI Standards and Innovation, the small Commerce Department unit that has been quietly reviewing frontier AI models, has been told to stop publishing what it finds. According to Gizmodo summarizing Wall Street Journal reporting, National Cyber Director Sean Cairncross was among the officials who instructed CAISI to pause public reports while a new presidential executive order is implemented.

The order in question, which Trump signed in June 2026, sets up a voluntary 30-day federal review window before frontier AI models reach public release, framed as a way to strengthen the cybersecurity of critical infrastructure. An earlier draft of that order reportedly ran longer, 90 days. The trigger for the public-reports freeze, per the reporting, was concern about what advanced systems, with Anthropic's Mythos cited by name, could do in the hands of someone trying to mount a cyberattack or build a biological weapon.

If you read this as a safety crackdown, the shape of it gets stranger. The unit being silenced is the one actually doing the safety work. CAISI was set up under the Biden administration and has been the government's main interlocutor with frontier labs; OpenAI recently called for strengthening the unit rather than restraining it. The administration's line, via White House spokeswoman Liz Huston, is that the implementation of President Trump's AI agenda is a whole-of-government effort. Some officials reportedly think the new executive-order regime simply duplicates what CAISI was already doing.

The honest caveat is that the directive doesn't shut CAISI down. Internal evaluations and inter-agency work continue, the voluntary agreements with AI developers remain in place, and nothing new becomes mandatory. What changes is who gets to see the findings. The reporting doesn't tell us what CAISI's unpublished reviews of Mythos and its peers actually concluded, whether the pause has an end date, or how much of Cairncross's office is angling for a bigger seat at the model-evaluation table.

For anyone working in frontier-model policy or building products on top of these systems, the public record of what they can do is now thinner exactly as the systems get more capable. The opening, for Congress, for academic red-teamers, and for the labs themselves, is to push for a reporting channel that survives any single executive order.