[UPDATE] White House Pivots Toward Frontier AI Oversight After Mythos

Three weeks after the White House postponed an AI safety executive order under industry pressure, the Trump administration is openly studying the kind of oversight it had previously rejected. The pivot is being driven by concerns about the national-security implications of Anthropic's Claude Mythos model, which the company released into a limited deployment under its Project Glasswing cybersecurity initiative, according to Fortune.

The Center for AI Standards and Innovation, the rebranded NIST-affiliated body funded by Congress in January, has signed agreements with Google DeepMind, Microsoft and Elon Musk's xAI to perform pre-deployment evaluations of frontier models. Existing testing partnerships with Anthropic and OpenAI have been renegotiated under the new mandate, CNBC reports.

What Mythos changed

Mythos was the inflection point. The model demonstrated a substantial capability lift in identifying and exploiting software vulnerabilities, which Anthropic described as a defensive tool for cybersecurity teams but which administration officials viewed as a category of risk worth gating. Seoul Economic Daily describes the conversation inside the White House as having shifted decisively after a classified briefing on the model's offensive potential.

The proposed framework would borrow language and structure from the U.S. Food and Drug Administration, with a working group of tech executives and government officials advising on what should count as a high-risk frontier model. Pre-release evaluation would be the gating step. The Federal News Network says the working group is meeting in closed sessions through June.

From anything goes to pre-release vetting

For the first months of the administration, the official line was that Biden-era AI safety efforts were a regulatory drag on American competitiveness. The reversal is conspicuous and has been noted across both the trade press and policy commentary. The Register characterised the move as a jump from "anything goes" to "strict regulation" inside a single news cycle.

States are not waiting. Colorado's comprehensive AI law remains scheduled to take effect on June 30, and bills are advancing in Washington, Florida, Virginia and Utah even as the December 2025 federal pre-emption order pushes the other way. Whether the new working group produces an actual executive order, and how it interacts with the state patchwork, will be the story to watch in the coming weeks.