Trump Scraps AI Executive Order at Last Minute, Citing Fears of Overregulation

On Thursday afternoon, with invitations already sent and staff assembled, the White House abruptly canceled the signing ceremony for what would have been the Trump administration’s most consequential AI policy action to date. President Trump told reporters the postponement was simple: “I didn’t like certain aspects of it.” Behind that terse explanation lies a months-long internal battle that exposes the fundamental tension at the heart of Washington’s approach to artificial intelligence—how to secure a technology while refusing to slow it down.

What the Order Would Have Done

The executive order had two major components. The first was a broad cybersecurity package: hardening Pentagon networks, launching an emergency hiring surge at the Cybersecurity and Infrastructure Security Agency (CISA)—an agency that had seen significant staffing cuts in recent years—and establishing mandatory threat-sharing mechanisms between private AI companies and the federal government whenever breaches occurred.

The second component was more unprecedented and more contentious. It would have established a “voluntary framework” under which frontier AI labs—companies developing the most capable models—would be required to share their models with the U.S. government at least 90 days before any public release. The National Security Agency would serve as the primary mediator for reviewing these models, with the authority to identify security vulnerabilities before they reached the public or adversaries. The framework also extended access rights to certain critical infrastructure operators, recognizing that utility grids, financial systems, and communications networks are increasingly dependent on AI.

On paper, the arrangement had buy-in from major players. Microsoft, Google, and Elon Musk’s xAI had already signed similar voluntary agreements with the government in early May, committing to pre-deployment testing. OpenAI and Anthropic were already operating under comparable arrangements through separate channels.

The Internal Battle That Killed It

According to reporting from Axios and the Washington Post, the order’s collapse was not a sudden presidential whim but the final outcome of weeks of fractious internal debate. Two factions within the White House had grown increasingly irreconcilable.

The pro-innovation camp—aligned with the administration’s broader deregulatory instincts—insisted the framework had to be genuinely voluntary, with minimal teeth and no real enforcement mechanism. For them, anything that created even a theoretical pathway to delayed AI releases was a threat to America’s competitive position against China. They worried that a 90-day pre-deployment window, even nominally voluntary, would eventually be used as leverage by regulators to extract compliance or delay products.

The national security bloc argued the opposite. With AI models increasingly capable of autonomously identifying and exploiting software vulnerabilities—a capability publicly demonstrated by Anthropic’s recently released “Mythos” model—allowing frontier labs to ship without any government review was a genuine national security risk. They wanted the NSA review to carry real weight, with classified testing protocols and actual enforcement backing the “voluntary” label.

Trump sided with neither faction cleanly. His public comments reflected both concerns simultaneously: praising AI as “causing tremendous good” while insisting the U.S. “is ahead of China” and he “doesn’t want to do anything that’s going to get in the way of that lead.”

The Anthropic Mythos Catalyst

Beneath the political maneuvering lies a specific technological trigger that made the executive order urgent in the first place. Anthropic’s Mythos model, released earlier this year, demonstrated an unsettling capability: given access to a codebase, it could autonomously scan for security vulnerabilities and generate functional exploits. In controlled testing environments, researchers showed the model could move from initial access to weaponizable attack code faster than any human red team.

That capability—what security researchers call “agentic cyber offense”—shifted the AI safety debate from abstract philosophical territory into concrete national security risk. The NSA and CISA began briefing Congressional staffers on implications. The executive order was, in part, a direct governmental response to the Mythos demonstration.

The irony is that Anthropic, the company whose model most directly motivated the policy, operates under existing voluntary agreements with the government already. The executive order was largely aimed at bringing the broader industry into a similar structure. Its collapse means that framework remains informal and inconsistent.

What Happens Next

The postponement creates immediate uncertainty across the AI industry. Labs that had been told the 90-day pre-deployment window was coming and had begun incorporating it into release planning now have no clarity on timing. Legal teams at major AI companies are in a holding pattern.

Administration officials told reporters Thursday that the order would be revised and could return “within weeks,” though no new signing date was specified. The cybersecurity components—CISA hiring, network hardening, threat sharing—are widely expected to survive any revision. The frontier model review framework is far less certain.

Meanwhile, the gap between AI capability and regulatory structure continues to widen. The government’s existing tools for managing advanced AI risks were designed for a world where models were powerful but passive—tools that humans controlled. The emergence of models with autonomous offensive cyber capabilities has outpaced every policy framework currently on the table.

Whether Trump signs a revised version, abandons the frontier model components entirely, or delegates the issue to Congress remains an open question. What’s clear is that Thursday’s last-minute cancellation was not a policy victory for any side—it was a demonstration of how poorly equipped Washington’s governance machinery is to keep pace with the technology it’s trying to oversee.

The White House did not provide a revised timeline as of publication.