US Weighs 30-Day Safety Reviews for Open-Source AI
The White House is considering broadening the US artificial-intelligence safety review framework beyond closed-source frontier systems to open-source models with comparable capabilities. The issue carries national-security and technology-governance implications while potentially reshaping competition in the US AI industry. Policymakers are seeking to avoid a system that subjects only proprietary developers to pre-release testing, which could impose unequal regulatory burdens on companies pursuing different development and distribution models.
Under the proposal described in the report, capability — rather than whether a model is open or closed — would determine whether it faces testing before release. Qualifying open-source AI models could be reviewed for as long as 30 days, matching the mechanism contemplated for closed frontier models. The White House aims to address security risks without weakening US competitiveness in open-source AI, though the report did not specify an effective date, the precise capability threshold or which agency would conduct the tests.
All Coverage
1 original reportsThe Backstory
The history behind this eventWhite House Exempts Open-Weight AI Models From Safety Tests
President Donald Trump’s June 2, 2026 executive order created a voluntary federal review for the most capable AI systems before wider release, giving agencies up to 30 days to assess cybersecurity and national-security risks. The Treasury Department, National Security Agency, Cybersecurity and Infrastructure Security Agency and National Institute of Standards and Technology were tasked with developing classified benchmarks. The policy matters because Washington is trying to manage frontier-model risks without slowing U.S. developers in the AI race with China.
At an August 4 closed-door meeting, White House officials told staff from Meta, Anthropic, Google, Nvidia and OpenAI that open-weight models would fall outside the unpublished framework, according to people familiar with the discussions. Only closed, proprietary U.S. models with state-of-the-art cyber capabilities and national-security risks would be covered. The carveout includes downloadable models from Chinese developers, undercutting Anthropic’s call for mandatory testing of any sufficiently capable system and prompting criticism over transparency, uneven oversight and a potential safety gap.
Subscribe to Mark Radar Weekly
Every Friday, the week's strongest signals in your inbox. Unsubscribe anytime.
If you search news on Google, you can set Mark Radar as a preferred source—our coverage will show up more often in your results. Set as preferred source on Google →