U.S. Presses Meta to Submit AI Models for Review

Reuters reports the Trump administration is pressing Meta to submit its AI models for voluntary review, citing a New York Times report quoting four people familiar with the confidential request, which was made in emails with the company. Reuters reports Meta is the only major U.S. developer that has not agreed to voluntarily share models with the federal government for review. Meta's emailed statement said: "We share the administration's goal of advancing U.S. leadership on robust and secure frontier AI. While we are working through the details, we hope to sign the agreement soon."

President Trump signed an executive order on June 2, 2026, establishing a voluntary framework for AI developers to submit "covered frontier models" to the U.S. government for up to 30 days before wider release. The original draft allowed up to 90 days, cut to 30 days in the final order after White House concerns about stifling innovation, per NPR. The order explicitly states it does not authorize mandatory licensing, preclearance, or permitting requirements; any binding regulation would require Congress.

Reuters reports that OpenAI and Anthropic were already working with the U.S. government to test unreleased models, while Google DeepMind, Microsoft, and xAI provide early access for national-security evaluations. The review process is administered through the Center for AI Standards and Innovation (CAISI). Meta remains the lone holdout among major U.S. frontier AI developers.

Voluntary pre-release reviews create a short window for government red-team exercises and vulnerability assessment of frontier models before broad deployment, targeting potential misuse vectors such as cybersecurity exploitation and dual-use military applications. For practitioners, the trend increases the practical importance of reproducible evaluation pipelines, pre-release risk documentation, and secure evidence-sharing workflows - and for open-weight developers specifically, it raises a harder question of how a pre-release review interacts with a release model that assumes the weights will be redistributed immediately and uncontrollably.

Whether a formal Meta-government agreement gets reported, and on what terms; how "covered frontier model" gets defined in practice, especially for open-weight releases; whether review scope expands beyond the current five labs; and whether technical test protocols get published that would define what a 30-day review can realistically evaluate for a model whose weights are meant to be downloaded rather than served behind an API.