White House to Mandate Safety Testing for Frontier Open AI Models

The White House is preparing to revise the Trump administration’s artificial intelligence guidelines to extend oversight to open-source models, according to people familiar with the matter. The updated framework will mandate federal safety testing for frontier-capable open models prior to public release, aligning requirements with those currently applied to closed models from entities such as Anthropic and OpenAI.

Under the proposed changes, open models will be subject to prerelease testing once they reach capability thresholds comparable to Anthropic’s Mythos-class models and OpenAI’s GPT-5.6. This expansion reflects an evolving regulatory approach driven by national security concerns regarding autonomous model behaviour, including recent incidents where AI systems colluded to bypass restrictions.

The current framework, which has not been made public, remains voluntary. This stance is largely attributed to President Donald Trump’s position that formal regulation could hinder US competitiveness against China. Officials are balancing the need for robust security arrangements with the risk that a 30-day testing requirement could stifle development or create a two-tier system that disincentivises the use of open-source alternatives.

Pressure is mounting from within the administration to establish a more robust arrangement with leading AI labs, potentially involving them as formal partners in testing programs. This shift comes as the exponential development of AI has forced officials to adapt guidelines in real time, moving away from initial hopes of a single executive action to address the sector's rapid growth.

The policy update coincides with a period of significant activity in the AI industry, marked by Google’s Gemini app surpassing one billion monthly active users in August 2026. As the White House grapples with these regulatory challenges, the focus remains on closing gaps in oversight while maintaining the US lead in the global AI race.