Trump White House Prepares Frontier AI Executive Order as Internal Divisions Slow Response to Mythos Security Threat

Overview

The Trump White House is preparing to release an executive order on AI safety as soon as this week — a striking reversal for an administration that began its term by dismantling Biden-era AI oversight. The order, according to Axios, would establish a voluntary framework under which AI labs share their frontier models with the government at least 90 days before public release. The shift has been forced by alarm over Anthropic’s Mythos model and OpenAI’s GPT-5.5-Cyber — systems whose “ability to find and exploit software vulnerabilities with unprecedented speed” spooked national security officials. But the path to the order has been slowed by internal disagreements that have exposed fault lines between the administration’s deregulatory instincts and its national security imperatives.

What We Know

The Executive Order’s Structure

The draft order contains at least two sections. The cybersecurity component aims to secure the Pentagon and other national security agencies, boost cyber hiring, strengthen hospital and bank systems, and encourage industry-government threat sharing about breaches. The frontier model component would involve multiple layers of government review to determine what qualifies as a “covered frontier model,” and then assess such models prior to their public release, Axios reports.

The proposed voluntary framework would require AI labs to share models with the government at least 90 days before public release and give access to certain critical infrastructure providers. The White House has not officially confirmed the order: a White House official said “any policy announcement will come directly from the President” and that “discussion about potential executive orders is speculation.”

The Catalyst: Mythos and GPT-5.5-Cyber

The proximate cause of the policy reversal is Anthropic’s Mythos model, which demonstrated an ability to identify decades-old software vulnerabilities that alarmed government cybersecurity officials. Platformer and Fortune report that Mythos — capable of identifying and exploiting cybersecurity vulnerabilities — triggered national security concerns that drove the policy reversal. The White House also moved to restrict Mythos access, opposing plans to expand the model’s availability from 50 to 120 companies.

The Anthropic situation has been particularly fraught. The Trump administration earlier designated Anthropic as a “supply chain risk” and barred federal agencies from using its technology — while simultaneously seeking to build evaluation frameworks around its most capable model.

The Ideological Reversal

The speed of the reversal is significant. On the first day of Trump’s second term, his administration revoked Biden’s executive order on AI safety testing. Three days later, Trump signed “Removing Barriers to American Leadership in Artificial Intelligence.” In February 2025, Vice President JD Vance told the Paris AI Action Summit that the “AI future is not won by hand-wringing about safety.” David Sacks, then the White House AI and crypto czar, dismissed critics as a “doomer industrial complex” engaged in “regulatory capture strategy.”

By May 2026, the same administration was arranging government pre-deployment reviews of the very models its officials once said needed no oversight.

CAISI: The Institution Built to Do the Work

The body tasked with pre-deployment evaluations is the Center for AI Standards and Innovation (CAISI) — the renamed Biden-era US AI Safety Institute, rebranded in June 2025. Under Chris Fall, a former Energy Department official and MITRE vice president, CAISI has already completed over 40 evaluations including of unreleased state-of-the-art models. In May 2026, Google DeepMind, Microsoft, and xAI agreed to give CAISI access to their models before public release, extending prior agreements with OpenAI and Anthropic.

But CAISI is under-resourced for the scope of what is being asked of it. The body has approximately 30 staff members and has received $30 million since its establishment in 2024, according to Federal News Network. Congress approved $10 million to expand CAISI in January 2026. The America First Policy Institute recommends $50–100 million in annual funding; the Federation of American Scientists proposes up to $155 million annually.

Internal Divisions Slow the Response

Progress toward the executive order has been disrupted by infighting. Axios reported on May 13 that National Economic Council Director Kevin Hassett had floated an “FDA for AI” approach, comparing future model evaluations to the process by which drug regulators ensure safety before market entry. Hassett stated the administration is “studying possibly an executive order to give a clear road map to everybody about how this is going to go and how future AIs that also could potentially create vulnerabilities should go through a process so that they’re released to the wild after they’ve been proven safe — just like an FDA drug.”

Chief of Staff Susie Wiles and Sacks subsequently walked back those comments. A Commerce Department webpage featuring frontier AI testing deals was also removed without explanation. A separate dispute has centered on whether pre-deployment testing authority should sit with the civilian Commerce Department or with national security agencies — a division that reflects deeper disagreements about whether frontier AI is primarily an economic asset or a national security threat.

According to Axios, disagreement among administration officials and scheduling pressures around President Trump’s China summit have further slowed efforts to finalize federal AI guidance. Sources told the outlet that momentum for the guidance “could run out quick.”

What We Don’t Know

Several critical details remain unresolved. It is unclear whether the “voluntary framework” in the draft will be strengthened to include any mandatory testing requirements, or whether it will remain a soft commitment susceptible to industry resistance. The draft’s definition of “covered frontier model” — the threshold that would trigger government review — has not been made public, leaving open questions about whether it would capture only the most powerful systems or a broader range of models.

It is also unclear which agency will ultimately hold authority over pre-deployment reviews. The civilian-versus-national-security debate within the administration remains unresolved, and the removal of the CAISI webpage suggests that institutional arrangements are still contested. Whether CAISI will receive the funding advocates say it needs to carry out the work at scale remains an open question for Congress.

Analysis

The pattern visible in Washington in mid-May 2026 is one of a deregulatory administration being overtaken by events it helped accelerate. By removing Biden-era oversight structures and signaling to industry that safety reviews were barriers rather than safeguards, the Trump administration helped create the conditions for the rapid development of the very capabilities now alarming its own national security apparatus. The Mythos episode has exposed a gap at the center of the administration’s AI posture: it wants American AI to win globally, but winning requires deploying capabilities that its own officials describe as serious security risks.

The institutional response — a renamed safety institute, voluntary industry agreements, a draft executive order — represents an improvised course correction rather than a coherent regulatory framework. Whether it coheres into durable policy depends on whether the administration can reconcile its internal divisions before the next frontier model arrives and resets the clock again.