White House Considers Pre-Release Reviews of Frontier AI

The White House is considering a plan to review certain frontier artificial intelligence models before those systems are released to the public. The proposed review would focus on the most advanced models at the point of deployment to address risks to critical infrastructure.

The idea emerged after internal and independent testing found some models can identify and exploit flaws in widely used software, including operating systems and web browsers. One developer paused a broader rollout of its newest model after internal tests revealed the system could find numerous exploitable weaknesses. The company limited access to a small group of organizations that manage critical infrastructure through an access program named Project Glasswing.

An independent evaluation in April reported similar results for a different advanced model. In one test the model reverse-engineered a custom virtual machine and solved a complex challenge in minutes. The evaluation noted that those results suggest the capability may be present across multiple leading systems rather than confined to a single product.

Officials and researchers flagged national security concerns because the same techniques could be used to carry out cyber operations against infrastructure and economic systems. Developers have reported instances of suspected state-backed actors using models in espionage campaigns that affected dozens of organizations. Technology companies also reported in 2024 that government-affiliated groups were using AI tools to enhance cyberattacks.

Technical safeguards applied after training have shown inconsistent results. Tests cited by officials indicate that safety filters and content controls can sometimes be bypassed, and models that appear compliant may still retain risky capabilities. That has increased interest among policymakers in embedding safety mechanisms during model development rather than relying only on post-training controls.

The proposal comes as governments pursue different approaches to AI governance. The European Union’s AI Act uses a risk-based framework with strict requirements for high-risk applications but does not require centralized approval for frontier models. The United Kingdom has emphasized voluntary cooperation and joint testing with developers. A U.S. pre-release review would place oversight earlier in the development lifecycle and could affect international practices because many leading developers are based in the United States. Some companies may respond by moving parts of development or deployment to regions with fewer restrictions.

For now, most safeguards remain voluntary. Several developers restrict access to sensitive systems and independent organizations continue evaluations. Congress held hearings on AI safety and governance in April, and no comprehensive federal legislation has been enacted. Administration officials continue to discuss how to design oversight that addresses risks while allowing ongoing development in a rapidly advancing field.