Trump seeks review of frontier models

President Donald Trump’s administration is preparing an executive action that would ask technology companies to submit some of their most advanced artificial intelligence models for review by federal agencies before public release, according to Politico and Reuters. The draft under discussion would extend Washington’s focus beyond AI-generated deepfakes and other downstream harms to the models themselves, people familiar with the matter told those outlets. Reuters reported the framework is expected to be voluntary, not mandatory, and could require developers to provide covered models to the government 90 days before launch. A White House spokesperson told Reuters that discussion of specific policy details was “speculation.” ### Which models are officials trying to reach before they go public? Politico reported the draft targets “advanced AI models,” part of a broader internal debate over how the federal government should handle frontier systems with national security implications. Reuters said the order under discussion would apply to “covered models” and would ask developers to engage with the U.S. government before public release. (politico.com) Kevin Hassett, director of the National Economic Council, said on Fox Business on May 6 that the administration was studying “possibly an executive order” so future AI systems that “potentially create vulnerabilities should go through a process” before they are released. Hassett compared the approach to how the Food and Drug Administration evaluates drugs for safety, according to Federal News Network. (politico.com) ### Why did this move gain traction inside the White House now? Anthropic’s “Mythos” model became a focal point in the administration’s internal debate, according to Reuters and Federal News Network. Federal News Network reported Anthropic previewed last month that Mythos could quickly find and exploit old software vulnerabilities, raising concerns that attackers could use AI to identify new flaws and build exploits faster than defenders can respond. (federalnewsnetwork.com) Sean Cairncross, the national cyber director, has been coordinating the government’s response, Hassett said in the Fox Business interview. Hassett said the administration had “scrambled an all-of-government effort” with the private sector to make sure the model was “tested left and right” before broad release. ### How far has the government already gone without a new order? (srnnews.com) The Commerce Department’s Center for AI Standards and Innovation, or CAISI, already has a voluntary role in testing frontier models. NIST says CAISI is tasked with establishing voluntary agreements with private-sector AI developers and leading evaluations of AI capabilities that may pose national security risks, including cybersecurity, biosecurity and chemical weapons concerns. (federalnewsnetwork.com) Federal News Network reported on May 6 that CAISI had announced new agreements with Google DeepMind, Microsoft and xAI to conduct pre-deployment evaluations of their frontier models. The outlet said CAISI already had similar agreements with Anthropic and OpenAI, and had conducted 40 evaluations, including on some unreleased models. Politico separately reported on May 5 that Microsoft, xAI and Google DeepMind had signed deals allowing the government to vet models ahead of release. (nist.gov) ### Who is pushing for tougher oversight, and who is resisting it? Reuters reported that some of Trump’s populist allies, including former adviser Steve Bannon and organizer Amy Kremer, have pressed the White House to require security testing for the most capable AI systems. The same report said venture capitalist Marc Andreessen and former Trump adviser David Sacks have opposed mandatory requirements. (federalnewsnetwork.com) David Sacks stepped down in March from his role as Trump’s lead AI official and is now co-chairing the president’s tech advisory committee, Reuters reported. The resulting fight has centered on whether the administration should preserve a light-touch approach to AI or impose a more formal pre-release process for top-end systems. (srnnews.com) ### How would this fit with Trump’s broader AI policy? The White House’s July 2025 “America’s AI Action Plan” emphasized accelerating AI innovation, encouraging open-source and open-weight AI, investing in interpretability and robustness, and building an AI evaluations ecosystem. The document also framed AI leadership as a national security priority. (srnnews.com) Any new order would land on top of that existing framework rather than replace it. Reuters reported Trump could sign an executive order on AI and cybersecurity as soon as Thursday, May 21, while Politico described the frontier-model review language as part of a still-evolving draft. The next concrete step is the text of any White House order and which agencies it names to receive model access before release. (srnnews.com) (whitehouse.gov)