Hours before tech CEOs were due to gather in the Oval Office on Thursday, President Trump called the ceremony off. The executive order on his desk would have created a voluntary framework for the federal government to review the most capable AI systems for national security risks before public release. He said he was worried it would slow the United States down. “We’re leading China, we’re leading everybody, and I don’t want to do anything that’s going to get in the way of that lead,” he told reporters.
The cancelled order is the visible part of a much messier debate inside the administration, and a useful window into how government is, or is not, adjusting to a generation of AI models that look less like chatbots and more like cybersecurity tools. According to reporting in the Los Angeles Times and Times Now, the immediate trigger was Anthropic’s Mythos model, which Treasury Secretary Scott Bessent and outgoing Fed chair Jerome Powell flagged to Wall Street CEOs in an urgent April meeting. Mythos, Anthropic has said publicly, can identify severe software vulnerabilities across major operating systems and browsers, and in some tests has helped relatively inexperienced researchers find remote code execution flaws overnight.
That capability is exactly why Anthropic restricted Mythos to a small group of partners under a programme it calls Project Glasswing, with Amazon Web Services, Microsoft, Google and CrowdStrike given limited access to harden their systems before similar capabilities show up in the wild. It is also what made Treasury nervous enough to convene the banks, and the White House nervous enough to draft an executive order in the first place. The draft, according to people familiar with it, would have given agencies including the Office of the National Cyber Director up to two months to design a review process under which firms shared frontier models 14 to 90 days before launch.
Anthropic and OpenAI were reportedly in talks with the White House about how such a system would work in practice. Both companies have plenty of reason to negotiate the terms of any review, given that the alternative is state level patchwork regulation or, eventually, congressional action they would have even less influence over. The administration has so far pushed back against state AI laws on the grounds they slow innovation. An executive order that effectively asked Washington to screen commercial models would have been a significant reversal of that posture.
There is also a parallel track that did not get cancelled. Earlier this month, the Commerce Department quietly announced agreements with Google, Microsoft and Elon Musk’s xAI to evaluate their most powerful models before public release, building on similar arrangements the Biden administration struck with Anthropic and OpenAI. The announcement was later pulled from the Commerce site. Whatever framework the administration eventually settles on, in other words, may already exist in fragments, just without a unifying signature.
Mythos itself complicates the politics in another way. In February, Trump ordered all US agencies to stop using Anthropic’s Claude chatbot after a public clash between the Pentagon and CEO Dario Amodei. The same government now considering whether to bring Anthropic’s most powerful model into a formal review process is in active legal conflict with the company. As Brown University’s Serena Booth put it to the AP, what looks like indecision from the outside is really fractures inside the coalition: a national security wing that wants more oversight, an innovation wing that wants less, and a White House that is not sure which to fight harder.
The deeper question, and the one Thursday’s non signing leaves unanswered, is whether the United States believes its frontier labs can self regulate at the level of capability Mythos represents. Anthropic appears to think the answer is yes, with caveats. Treasury and parts of the national security establishment appear to think the answer is no. For now the model is in the hands of a small group of partners, the executive order is in a drawer, and the gap between what the most capable AI systems can do and what the government is permitted to know about them remains roughly where it was a month ago.