Timeline

White House blocks Anthropic from expanding Mythos access, weighs pre-release vetting regime

Officials cited leak risk and worry the NSA's compute share would shrink, while separately telling Anthropic, Google and OpenAI they were weighing government review of models before release.

  • Government & policy
  • Notable

The Trump administration rejected Anthropic’s request to roughly double, from about 50 to 120 organisations, the partner list for Project Glasswing, the restricted-access programme through which Anthropic had been sharing its Mythos Preview model — withheld from public release over its ability to autonomously find and exploit zero-day vulnerabilities. Officials cited two concerns: that access controls could not be trusted to prevent leaks at that scale, pointing to an April incident in which unauthorised users had already guessed a Mythos access URL, and that adding 70 more partners would crowd out the National Security Agency’s own share of the model’s compute, which was being used to hunt vulnerabilities in Microsoft software.

Days later, the New York Times reported that the White House was separately discussing a more sweeping step: an executive order establishing a working group to review advanced AI models before public release, a reversal of the administration’s 2025 rescission of Biden-era pre-deployment testing requirements. Officials had already briefed executives from Anthropic, Google and OpenAI on the plans, framed partly as protecting the administration from political fallout should an AI-assisted cyberattack occur, and partly as evaluating whether frontier models could be useful to the Pentagon and intelligence agencies. One model under consideration was Britain’s approach of assigning existing government bodies to check models against safety standards.

The two developments sat awkwardly together: the same week it froze Anthropic’s commercial expansion, the administration was also drafting a mechanism to give federal agencies their own access to Mythos regardless of the Department of War’s supply-chain-risk designation against the company. It was an early instance of the executive branch directly shaping which organisations could use a specific frontier model, ahead of the formal export-control action taken against Mythos and its consumer version, Fable, the following month.