White House to Expand AI Safety Testing to Open-Weight Models

The White House is expanding its artificial intelligence oversight framework to include open-weight models, ending a policy that previously exempted them from federal safety testing. This shift marks a significant pivot in how the U.S. government manages the risks associated with advanced AI development.

According to reports from WIRED and Axios, the administration is moving away from a system that focused primarily on proprietary models from companies like OpenAI and Anthropic. The initial framework, reported in August 2026, explicitly excluded open models from safety testing for advanced AI capabilities. However, the government is now transitioning toward a risk-based approach. Under this new direction, oversight will be determined by a model's demonstrated capabilities and the risks it poses, regardless of whether the model is open or closed.

The Tension of Open Innovation

This policy shift stems from an ongoing tension between the desire to promote open-source innovation and the need to mitigate national security risks. Early in the current administration's approach, open-weight models were viewed as essential for maintaining U.S. competitiveness and ensuring transparency. This perspective led to the initial decision to exempt these models from rigorous federal vetting, allowing developers to iterate quickly without the burden of government oversight.

As open models have grown more powerful, federal officials are reconsidering whether the "open" nature of a model justifies a lack of safety testing. The core of the debate is whether a model possessing dangerous capabilities should be subject to the same scrutiny as a corporate lab's product, even if its weights are publicly available.

Implications for the AI Ecosystem

This transition signals a move from a permissive environment for open AI to one where high-capability open models are viewed as potential security liabilities. For the open-source community, this could introduce significant regulatory hurdles. Testing requirements that were once the sole domain of giant corporate labs may now apply to smaller teams and independent researchers who release open-weight models.

By treating capabilities as the trigger for oversight rather than the business model, the White House is effectively narrowing the gap between proprietary and open-source regulation. This could slow the pace of open-source releases if the cost and time associated with federal safety testing become prohibitive for non-corporate entities.

What to Watch

As the White House continues to grapple with how to regulate a technology it has previously tried to leave alone, the industry is waiting for the specific triggers that will mandate testing. While the shift toward a risk-based framework is confirmed, the exact thresholds for what constitutes a "risk" or a "capability" sufficient to trigger federal intervention remain undefined. Future updates to the framework will likely detail the specific cybersecurity benchmarks and safety tests that open-weight developers must now pass.