The Trump administration is reportedly weighing an executive order to create government oversight of new AI models before public release, a sharp reversal from its earlier deregulatory stance and a potential shift toward formal pre-clearance that would reshape the competitive dynamics of frontier AI labs.
The New York Times reported that senior officials briefed executives from Anthropic, Google, and OpenAI on plans for an AI working group of tech leaders and government officials to examine review procedures for new models. The discussions include a systematic government review process similar to one being developed in the United Kingdom, where agencies ensure AI models meet safety standards. The change appears prompted by concerns about Anthropic's Mythos model, which cybersecurity experts warn could supercharge complex cyberattacks through unprecedented vulnerability discovery and exploitation. White House officials declined to confirm the report, saying any policy announcement would come directly from the president.
For OpenAI, Anthropic, Google DeepMind, Meta, and xAI, pre-release review would mean formal government evaluation before model deployment. The largest labs have already signed voluntary safety testing agreements with the government, but those are cooperative arrangements without mandatory pre-clearance. A formal review process would create delays, documentation requirements, and compliance costs that affect release cadence. OpenAI and Anthropic release models every six to nine months. Google DeepMind and Meta operate on similar timelines. xAI moves faster but is smaller. A 90-day government review would compress the window between research breakthrough and market deployment, favouring labs with parallel development pipelines and regulatory teams.
The incumbent advantage is structural. OpenAI and Anthropic have dedicated safety and policy teams that already engage with government on voluntary testing. Google DeepMind and Meta have regulatory affairs groups with decades of experience in Washington. Those labs can absorb the overhead of pre-clearance without slowing innovation. Smaller startups and open-source projects lack the resources for mandatory reviews, which could entrench the frontier labs as the only players capable of shipping new models at scale. The review process would also create a bifurcated market where government-approved models have a trust advantage for enterprise customers, while unapproved models face adoption barriers.
Washington's shift from voluntary access to formal control is quiet but real. The Biden administration's AI executive order in 2023 required safety testing reports for models with extreme risk potential, but it was light-touch and focused on transparency. The Trump administration started deregulatory, with David Sacks and Marc Andreessen advising against heavy-handed rules. The Mythos incident appears to have changed that calculus. Cybersecurity risks from models that can autonomously discover and exploit vulnerabilities are concrete and immediate. The White House is now contemplating technical guidelines for securing open-weight models and tapping the intelligence community for oversight. That is a move from cooperation to control.
For SF founders, the implications are straightforward. Release cadence is now a core competitive advantage, and government review would turn speed into a privilege that only incumbents can afford. Startups would face higher compliance costs, longer timelines to market, and a trust disadvantage when selling to regulated industries. Fundraising narratives would need to account for regulatory risk, and partnerships with frontier labs would become more attractive as a way to piggyback on their approved models. The open-source ecosystem would suffer most, as community models lack the resources for formal review and the incentive to seek it.
The policy shift also creates opportunities for founders who build compliance and safety infrastructure. Tools for automated safety testing, documentation generation, and audit trails would become table stakes for any lab seeking government approval. Consulting firms specialising in AI regulatory navigation would thrive. Platforms that aggregate approved models for enterprise deployment would capture value. The review regime would not kill innovation, but it would channel it toward companies that can navigate the process.
Also read: Agent startups are chasing the wrong moat, and the market is already separating demos from durable businesses • Barry Diller trusts Sam Altman, but says trust is the wrong tool for governing AGI • ZAYA1-8B is an AMD-trained small model that tests whether frontier intelligence can escape Nvidia's CUDA gravity
