The Trump administration spent three weeks signaling that it might require government approval before AI companies could release their most powerful models. According to Bloomberg, the executive order being prepared will do no such thing.

The draft order, reported by Bloomberg on May 8, will focus on protecting critical infrastructure and government networks from AI-enabled cyberattacks. It will create technical guidelines and best practices for securing open-weight models. What it will not do is require the government to vet frontier AI models before they are released to the public.

The gap between what was floated and what is being delivered is significant.

How We Got Here

The story starts with Anthropic's Mythos model. When the company previewed its capabilities last month, cybersecurity experts warned that Mythos could find and exploit decades-old software vulnerabilities faster than human hackers. The national security establishment panicked. The White House, which had revoked Biden's AI safety executive order on its first day in office, suddenly found itself considering something remarkably similar.

On May 4, The New York Times reported the administration was considering a formal government review process for new AI models. Kevin Hassett, then the top White House economic adviser, compared the potential system to the FDA's drug approval process. "We might look at AI models the way we look at pharmaceuticals," he told reporters.

In meetings, White House officials told executives from Anthropic, Google, and OpenAI about some of those plans. Google, Microsoft, and Elon Musk's xAI signed voluntary agreements allowing the Commerce Department to test their models before public release. Forbes reported on May 4 that the deals were framed as a way to avoid mandatory regulation.

What the EO Actually Does

The 16-page draft directive, according to people familiar with the matter, focuses on defensive measures. It would create guidelines for securing open-weight AI models, which have publicly available training parameters and can be adapted by anyone. It addresses AI-enabled cyberattack threats to government networks and critical infrastructure.

What it explicitly avoids: any requirement for government sign-off before companies release their most capable models. The FDA analogy is dead. The administration chose voluntary cooperation over mandatory oversight.

Industry Won. The Question Is What They Won.

This is a clear victory for the AI industry's lobbying apparatus. Seven lobbyists and advisers told Politico last week that they could not get specifics on the proposed EO, and the confusion was itself creating business uncertainty. The resolution, it turns out, is that the administration pulled back from the most aggressive version.

But the victory is ambiguous. The voluntary testing agreements with Google, Microsoft, and xAI are real. Commerce Department officials will have access to new models before public release. The difference is that the companies volunteered. There is no legal requirement and no penalty for opting out. Anthropic, notably, is not among the companies that signed.

Independent assessments have questioned whether the Mythos threat is as severe as Anthropic claims. Research from AISLE Security found that open-source models could detect many of the same flagship vulnerabilities that were presented as uniquely dangerous. The Pentagon has its own complicated relationship with Anthropic, having effectively blacklisted the company while still wanting access to Mythos for national security applications.

The Regulatory Void Remains

The irony of the timing is hard to miss. This EO will land during the same period that Europe watered down the EU AI Act after a nine-hour negotiation, that Connecticut passed America's most aggressive state-level AI law, and that Colorado pushed through its own algorithmic discrimination bill. The two largest economies on earth are moving in opposite directions, with no coherent framework emerging from either.

At the state level, the patchwork is accelerating. There are now over 1,200 AI bills across state legislatures, with 145 enacted. At the federal level, the story is the same as it has been for months: lots of rhetoric, some voluntary agreements, and no enforceable rules for the companies building the most powerful AI systems in history.

Mythos scared the White House into considering real oversight. The executive order shows how quickly that impulse fades when industry pushes back.

First reported by Bloomberg. Additional reporting from Politico, Axios, The Hill, Forbes, and Tom's Hardware.