President Trump pulled a long-drafted AI cybersecurity executive order on May 21, hours before a Rose Garden-style signing ceremony to which tech CEOs had already been invited. “I didn’t like certain aspects of it. I postponed it,” Trump said.

The two-section draft, first reported by Axios, would’ve required frontier AI labs to share models with the federal government on a voluntary basis at least 90 days before public release, with the Office of the National Cyber Director leading the evaluation track. A second section covered cybersecurity hiring and threat-sharing.

White House AI adviser David Sacks called the framework “something doomers wanted.” One source said Sacks “hated it”; another said Trump “just hates regulation.” TechCrunch reported a softer contributing reason: not enough CEOs could get to Washington on short notice for the photo op.

The substantive industry objection, per Axios, was the assignment of Treasury to a lead role in surfacing AI security flaws, work normally housed at CISA and NIST. Open questions also remained about whether the 90-day government hold would prevent labs from sharing models with allied governments for parallel testing.

The proximate policy driver is sitting in production. Anthropic says its Mythos model finds and exploits network vulnerabilities at unprecedented speed; the NSA is already using it, and the White House recently blocked Anthropic from distributing Mythos to dozens of additional firms over security concerns. The order was, in effect, an attempt to systematize what’s already happening on an ad hoc basis.

With the order shelved and no new signing date set, the NIST Center for AI Standards and Innovation’s voluntary testing program, which counts Anthropic, Google, Microsoft, xAI, and OpenAI as participants, remains the only federal pre-release review vehicle. That outcome resembles the post-2023 Biden order rollback more than anyone in the building probably wanted to admit.

Sources

Sources