President Donald Trump signed an executive order on Thursday creating a voluntary federal-access framework for frontier artificial-intelligence models, the most significant policy turn on AI safety from the second Trump administration. The order requires participating labs to share new models with the government 90 days ahead of public release, and to extend the same window to critical-infrastructure operators including major banks, utilities and telecoms.
The signing ceremony in the Oval Office included Sam Altman of OpenAI, Dario Amodei of Anthropic, Sundar Pichai of Alphabet, Satya Nadella of Microsoft, Jensen Huang of Nvidia and Mark Zuckerberg of Meta. The presence of all six firms was a turnaround from January, when the same administration tore up the cybersecurity testing regime its predecessor had layered on top of frontier models.
The trigger, according to senior officials, was the release in March of OpenAI's GPT-5.5-Cyber model and the rolling preview of Anthropic's "Mythos" system, both of which independent red teams found capable of automating significant portions of network-intrusion workflows. White House cybersecurity staff have spent the past six weeks lobbying internally for the new framework against opposition from market-oriented advisers.
The new order is structured as voluntary participation tied to federal procurement preference. Labs that opt in will receive expedited certification for sales into the Department of Defense, the intelligence community and the federal civilian agencies. Labs that opt out will not be excluded from those markets outright, but will face a longer review.
Among the conditions attached to participation, models exceeding a compute threshold initially set at 10^26 floating-point operations during training must be shared with a new "Cybersecurity Evaluation Cell" run jointly by CISA and the NSA. Models that score above an as-yet-undefined threshold on the cell's red-team battery may have their public release deferred by up to a further 60 days.
Anthropic and OpenAI both said in statements they would participate. Meta's position on its open-weights Llama family remains unclear; the company indicated it would "engage constructively" but stopped short of committing to share weights pre-release. Google said its Gemini frontier work was already being shared with US allies' AI safety institutes and that the new framework would formalise an existing practice.
The order leaves the European Union, the United Kingdom and Japan out of the pre-release loop, drawing sharp criticism from the AI Safety Institute network. Britain's technology secretary said London would seek a parallel bilateral arrangement with Washington. Brussels said the move was "inconsistent with the principle of co-regulation" but stopped short of suggesting retaliation.