White House infighting over AI cybersecurity oversight leaves major labs uncertain whether a revised executive order will emerge and what it would actually require.
Trump scrapped a planned executive order on artificial intelligence oversight on May 21, canceling the signing ceremony hours before it was set to take place. His explanation to reporters: the order risked blunting domestic AI competition and narrowing the US advantage over China.
The cancellation exposed a fractured White House. Wired reports that officials have spent weeks in chaotic internal talks about whether the order can be salvaged. AI executives from major labs say they remain unclear on what a revised version would require, or whether one will be signed at all.
The intelligence argument
At the heart of the draft order was a voluntary review framework: labs including OpenAI, Anthropic, and Google would submit AI models to the White House for cybersecurity assessment up to 90 days before public release. The rationale came from specific capability concerns. Officials pointed to Anthropic's Mythos and OpenAI's GPT-5.5 as evidence that frontier artificial intelligence has crossed into national security territory, given both models' demonstrated ability to find vulnerabilities in legacy software systems.
For the labs, a 90-day pre-release window is a substantial operational demand. Several executives told Wired their companies are not set up to share models that far ahead of time. Some aides and industry figures still hold out hope that a revised order could pass with its most contentious provisions stripped out.
Anthropic's parallel moves
The White House deliberations arrive at an awkward moment for Anthropic. Yahoo Finance reported Tuesday that the company confidentially filed IPO paperwork while simultaneously broadening the enterprise program built around the very models at the center of the regulatory debate. Project Glasswing now covers 150 companies, up from 50, giving paying customers tools to probe their own code for weaknesses using Mythos. Anthropic has itself previously flagged concerns about Mythos's ability to surface cybersecurity gaps in production systems.
That expansion will complicate any White House negotiation. If Mythos-level capabilities are already in the hands of 150 enterprise clients, the logic for a pre-release government review becomes harder to sustain.
A policy reversal stalled
The aborted order marked a genuine ideological shift for an administration that arrived in Washington broadly skeptical of AI regulation. Acknowledging that model development has outpaced the government's ability to assess risks in real time is, quietly, a significant concession.
As Forbes noted earlier this year, OpenAI alone now counts 800 million monthly active users and $20 billion in annual recurring revenue. At that scale, self-regulation arguments carry less weight than they once did. With Anthropic filing for an IPO and Google iterating models at speed, the major labs are entering a phase of commercial and institutional heft that makes the oversight question harder to defer.
A US policy deadlock does not eliminate rules; it means other jurisdictions write them first. The EU's AI Act is already in enforcement mode, and China has issued AI governance rules at a pace most observers underestimated. Companies operating globally navigate a patchwork regardless. Even specialized players in adjacent spaces are betting on regulatory complexity: Munich-based legal AI startup Bayshore raised $8 million this week to encode compliance rules into AI agents, according to Analytics Insight, on the premise that the gap between what law requires and what businesses can execute will keep widening.
Some White House officials still believe a narrowed order, targeting a defined set of capability thresholds for national security review, could move forward. Others think the moment has passed. The labs are not waiting to find out.
The question the administration keeps avoiding: who decides when an AI model is powerful enough to require government review before it reaches the market? Right now, the answer is nobody.
---
FAQ
What did Trump's canceled AI executive order propose?
The draft would have required AI labs including OpenAI, Anthropic, and Google to provide the White House with early access to new models, up to 90 days before public release, for cybersecurity evaluation under a voluntary framework.
Why did Trump cancel the AI regulation signing ceremony?
Trump said publicly that the order could stifle domestic AI competition and erode the US lead over China. Reports indicate internal White House disagreements also played a significant role in the last-minute cancellation.
What is Anthropic's Mythos model and why does it matter here?
Mythos is Anthropic's frontier AI model, notable for its ability to identify vulnerabilities in software systems. The administration cited it as one reason AI had become a national security concern; Anthropic has since expanded enterprise access through Project Glasswing.
Could a revised executive order on AI still be signed?
Some administration officials and industry figures believe a narrower version focused on specific national security thresholds remains possible, but multiple people familiar with the deliberations describe the process as chaotic with no clear timeline.